Add 21 accepted v2 seed31415 instances (81 tasks total)
#2
by panjose - opened
This view is limited to 50 files because it contains too many changes. See the raw diff here.
- .gitattributes +1 -0
- README.md +30 -2
- metadata.json +70 -5
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/data/experiments.json +1054 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/data/observations.npz +3 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/data/study.json +14 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/instance_meta.json +18 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/prompt_b1.md +67 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/prompt_b2.md +58 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/prompt_b3.md +42 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/prompt_b4.md +46 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/reference/ground_truth.npz +3 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/reference/predictions.npz +3 -0
- tasks/biology.mixed_perturbation_gene_transfer__seed31415/reference/report.md +1 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/data/promoters.fa +80 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/data/task_info.json +18 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/instance_meta.json +15 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/prompt_b1.md +63 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/prompt_b2.md +49 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/prompt_b3.md +34 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/prompt_b4.md +36 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/reference/motif_ref.json +81 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/reference/motif_ref.txt +1 -0
- tasks/biology.subtle_regulatory_motif_discovery__seed31415/reference/sites_ref.csv +41 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/data/README.md +1 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/data/genes.json +0 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/data/manifest.json +35 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/data/metadata.json +72 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/data/provenance.json +22 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/data/test_queries.json +241 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/data/train_condition_stats.npz +3 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/data/validation_targets.npz +3 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/instance_meta.json +23 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/prompt_b1.md +441 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/prompt_b2.md +379 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/prompt_b3.md +192 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/prompt_b4.md +196 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/README.md +1 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/genes.json +0 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/manifest.json +35 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/metadata.json +72 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/provenance.json +22 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/test_queries.json +241 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/train_condition_stats.npz +3 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/validation_targets.npz +3 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/reference.json +26 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/reference.npz +3 -0
- tasks/biology.unseen_perturbation_rna_response__seed31415/reference/results/predictions.npy +3 -0
- tasks/biostat.adaptive_cohort_transport__seed31415/data/experiments.json +0 -0
- tasks/biostat.adaptive_cohort_transport__seed31415/data/policies.json +0 -0
.gitattributes
CHANGED
|
@@ -3,3 +3,4 @@
|
|
| 3 |
*.h5 filter=lfs diff=lfs merge=lfs -text
|
| 4 |
*.hdf5 filter=lfs diff=lfs merge=lfs -text
|
| 5 |
*.png filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 3 |
*.h5 filter=lfs diff=lfs merge=lfs -text
|
| 4 |
*.hdf5 filter=lfs diff=lfs merge=lfs -text
|
| 5 |
*.png filter=lfs diff=lfs merge=lfs -text
|
| 6 |
+
tasks/computer_science.retry_storm_fault_localization__seed31415/data/traces.jsonl filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -24,7 +24,7 @@ task_categories:
|
|
| 24 |
|
| 25 |
## Dataset Summary
|
| 26 |
|
| 27 |
-
[ASI-Bench](https://github.com/apexin-ai/ASI-Bench) evaluates general intelligence, innovation, and autonomous execution through
|
| 28 |
|
| 29 |
This repository contains the fixed seed `31415` generated instances, including prompts, task inputs, reference artifacts, and available instance-generation metadata.
|
| 30 |
|
|
@@ -57,4 +57,32 @@ https://arxiv.org/abs/2608.17271
|
|
| 57 |
|
| 58 |
## License
|
| 59 |
|
| 60 |
-
This dataset is licensed under the [Apache License 2.0](./LICENSE).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
## Dataset Summary
|
| 26 |
|
| 27 |
+
[ASI-Bench](https://github.com/apexin-ai/ASI-Bench) evaluates general intelligence, innovation, and autonomous execution through 81 project-level scientific research tasks spanning 12 domains. Each task provides four matched prompt conditions: B1 gives full methodological guidance, B2 identifies the method without the full procedure, B3 provides only the research objective and data, and B4 adds task-irrelevant information to B3.
|
| 28 |
|
| 29 |
This repository contains the fixed seed `31415` generated instances, including prompts, task inputs, reference artifacts, and available instance-generation metadata.
|
| 30 |
|
|
|
|
| 57 |
|
| 58 |
## License
|
| 59 |
|
| 60 |
+
This dataset is licensed under the [Apache License 2.0](./LICENSE).
|
| 61 |
+
## Accepted v2 additions (2026-11-10)
|
| 62 |
+
|
| 63 |
+
This update adds 15 accepted tasks, one seed31415 instance each. The original 60
|
| 64 |
+
instances are unchanged. Prompt templates and numerical references live here;
|
| 65 |
+
generators and private solver source are excluded. The option-surface reference
|
| 66 |
+
contains a pre-materialized replay bundle consumed by the public scorer. These
|
| 67 |
+
new additions have not been included in the original 60-task reported results.
|
| 68 |
+
|
| 69 |
+
## Accepted additions (2026-11-17)
|
| 70 |
+
|
| 71 |
+
This PR adds five more seed31415 instances: QP sensitivity audit, minimal latent
|
| 72 |
+
realization, heterogeneous click-log counterfactuals, signed Ising log partition,
|
| 73 |
+
and A2 trinion full index. The sixth requested RNA response task is excluded
|
| 74 |
+
because its supplied ZIP is truncated. Together with the earlier 15 additions,
|
| 75 |
+
this PR contains 80 tasks; main remains unchanged until a dataset maintainer
|
| 76 |
+
merges it. New tasks are outside the original 60-task reported benchmark results.
|
| 77 |
+
Private generators and reference programs are excluded; numerical references
|
| 78 |
+
and public input problems needed by scoring are included.
|
| 79 |
+
|
| 80 |
+
## RNA upload completion (2026-11-17)
|
| 81 |
+
|
| 82 |
+
The intact RNA response archive has now been received and validated. This
|
| 83 |
+
supersedes the earlier pending/truncated-upload note: all six November 17 tasks
|
| 84 |
+
are included, bringing this PR to 81 tasks (60 original plus 21 accepted additions).
|
| 85 |
+
The RNA task retains its original fixed measured-data split; harness seed31415
|
| 86 |
+
does not create a new scientific split. Numerical references needed for local
|
| 87 |
+
seed31415 scoring are included; private source provenance, chunked generator
|
| 88 |
+
assets, GT generators and private solver programs are excluded.
|
metadata.json
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
{
|
| 2 |
-
"name": "ASI-Bench
|
| 3 |
"seed": 31415,
|
| 4 |
"seed_overrides": {
|
| 5 |
"computer_science.deployment_prediction_sets": 2,
|
|
@@ -8,8 +8,8 @@
|
|
| 8 |
"includes_reference_answers": true,
|
| 9 |
"source_repository": "apexin-ai/Agent-AI4Sci-Bench",
|
| 10 |
"source_commit": "6192234d4b9fdd8657a8054ac12b4cbaa827ea26",
|
| 11 |
-
"task_status": "
|
| 12 |
-
"task_count":
|
| 13 |
"task_ids": [
|
| 14 |
"astronomy.nbody_close_encounters",
|
| 15 |
"astronomy.planet_activity_sep",
|
|
@@ -18,8 +18,12 @@
|
|
| 18 |
"biology.contextual_dna_regime_mapping",
|
| 19 |
"biology.gene_network_timeseries",
|
| 20 |
"biology.gillespie_ssssa_michaelis_menten",
|
|
|
|
| 21 |
"biology.protein_folding_go_model",
|
| 22 |
"biology.scrna_seq_analysis",
|
|
|
|
|
|
|
|
|
|
| 23 |
"biostat.perturbation_convergence_testing",
|
| 24 |
"chemistry.implicit_solvent_pb_gb",
|
| 25 |
"chemistry.marcus_et_reorganization_energy",
|
|
@@ -33,15 +37,29 @@
|
|
| 33 |
"computer_science.deployment_prediction_sets",
|
| 34 |
"computer_science.dp_accountant_forensics",
|
| 35 |
"computer_science.euclidean_tsp_tour_optimization",
|
|
|
|
|
|
|
| 36 |
"computer_science.hmc_neal_funnel",
|
| 37 |
"computer_science.max3sat_assignment_optimization",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
"computer_science.stable_latent_state_fitting",
|
| 39 |
"earth_science.2_e_soil_carbon_nitrogen_cycling",
|
| 40 |
"earth_science.farquhar_ballberry_photosynthesis",
|
|
|
|
|
|
|
| 41 |
"earth_science.richards_topmodel_hydrology",
|
|
|
|
| 42 |
"earth_science.snow_3layer_compaction",
|
|
|
|
| 43 |
"electrical_engineering.cmos_opamp_design",
|
| 44 |
"electrical_engineering.lossy_tline_skin",
|
|
|
|
|
|
|
|
|
|
| 45 |
"materials.ising_domain_growth",
|
| 46 |
"materials.relaxation_mode_recovery",
|
| 47 |
"math.homotopy_poly_roots",
|
|
@@ -51,12 +69,14 @@
|
|
| 51 |
"math.mpsc_safety_filter",
|
| 52 |
"math.multiply_connected_conformal_mapping",
|
| 53 |
"math.nnls_modulus_deblur",
|
| 54 |
-
"math.ucb_q_learning_regret",
|
| 55 |
"math.sigma_linear_diophantine",
|
| 56 |
"math.sparse_hyperbolic_recovery_2d",
|
| 57 |
"math.sparse_nonlinear_interaction_recovery",
|
| 58 |
"math.symplectic_integrators",
|
|
|
|
|
|
|
| 59 |
"physics.ag_strip_dielectric_inference",
|
|
|
|
| 60 |
"physics.dp_contact_process",
|
| 61 |
"physics.ks_2d_anisotropic",
|
| 62 |
"physics.kt_phase_transition_2d_xy_model",
|
|
@@ -64,6 +84,7 @@
|
|
| 64 |
"physics.lid_driven_cavity_flow",
|
| 65 |
"physics.paired_nanorod_negative_index_retrieval",
|
| 66 |
"physics.vpfm_leapfrog",
|
|
|
|
| 67 |
"robotics.lidar_imu_time_offset_yaw_calibration",
|
| 68 |
"robotics.multiregime_icp_degeneracy",
|
| 69 |
"robotics.occupancy_grid_mapping_correlated_clutter",
|
|
@@ -74,5 +95,49 @@
|
|
| 74 |
],
|
| 75 |
"generated_at_utc": "2026-08-03T08:14:22Z",
|
| 76 |
"license": "Apache-2.0",
|
| 77 |
-
"license_file": "LICENSE"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 78 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"name": "ASI-Bench Generated Instances \u2014 Seed 31415",
|
| 3 |
"seed": 31415,
|
| 4 |
"seed_overrides": {
|
| 5 |
"computer_science.deployment_prediction_sets": 2,
|
|
|
|
| 8 |
"includes_reference_answers": true,
|
| 9 |
"source_repository": "apexin-ai/Agent-AI4Sci-Bench",
|
| 10 |
"source_commit": "6192234d4b9fdd8657a8054ac12b4cbaa827ea26",
|
| 11 |
+
"task_status": "60 original release tasks plus 21 accepted task additions (pending PR)",
|
| 12 |
+
"task_count": 81,
|
| 13 |
"task_ids": [
|
| 14 |
"astronomy.nbody_close_encounters",
|
| 15 |
"astronomy.planet_activity_sep",
|
|
|
|
| 18 |
"biology.contextual_dna_regime_mapping",
|
| 19 |
"biology.gene_network_timeseries",
|
| 20 |
"biology.gillespie_ssssa_michaelis_menten",
|
| 21 |
+
"biology.mixed_perturbation_gene_transfer",
|
| 22 |
"biology.protein_folding_go_model",
|
| 23 |
"biology.scrna_seq_analysis",
|
| 24 |
+
"biology.subtle_regulatory_motif_discovery",
|
| 25 |
+
"biology.unseen_perturbation_rna_response",
|
| 26 |
+
"biostat.adaptive_cohort_transport",
|
| 27 |
"biostat.perturbation_convergence_testing",
|
| 28 |
"chemistry.implicit_solvent_pb_gb",
|
| 29 |
"chemistry.marcus_et_reorganization_energy",
|
|
|
|
| 37 |
"computer_science.deployment_prediction_sets",
|
| 38 |
"computer_science.dp_accountant_forensics",
|
| 39 |
"computer_science.euclidean_tsp_tour_optimization",
|
| 40 |
+
"computer_science.generate_conforming_hexahedral_elements_with_planar_faces",
|
| 41 |
+
"computer_science.heterogeneous_click_log_counterfactuals",
|
| 42 |
"computer_science.hmc_neal_funnel",
|
| 43 |
"computer_science.max3sat_assignment_optimization",
|
| 44 |
+
"computer_science.minimal_latent_realization",
|
| 45 |
+
"computer_science.planted_clique_community_recovery",
|
| 46 |
+
"computer_science.qp_sensitivity_audit",
|
| 47 |
+
"computer_science.retry_storm_fault_localization",
|
| 48 |
+
"computer_science.signed_ising_log_partition",
|
| 49 |
"computer_science.stable_latent_state_fitting",
|
| 50 |
"earth_science.2_e_soil_carbon_nitrogen_cycling",
|
| 51 |
"earth_science.farquhar_ballberry_photosynthesis",
|
| 52 |
+
"earth_science.flood_peak_reconstruction",
|
| 53 |
+
"earth_science.hysteretic_snowmelt_runoff_forecast",
|
| 54 |
"earth_science.richards_topmodel_hydrology",
|
| 55 |
+
"earth_science.robust_model_inversion_of_insar_deformation_at_an_active_volcano",
|
| 56 |
"earth_science.snow_3layer_compaction",
|
| 57 |
+
"electrical_engineering.battery_transient_thermal_forecast",
|
| 58 |
"electrical_engineering.cmos_opamp_design",
|
| 59 |
"electrical_engineering.lossy_tline_skin",
|
| 60 |
+
"electrical_engineering.vector_pair_limits",
|
| 61 |
+
"finance.robust_arbitrage_free_option_surface",
|
| 62 |
+
"materials.fcc_tilt_gb_family",
|
| 63 |
"materials.ising_domain_growth",
|
| 64 |
"materials.relaxation_mode_recovery",
|
| 65 |
"math.homotopy_poly_roots",
|
|
|
|
| 69 |
"math.mpsc_safety_filter",
|
| 70 |
"math.multiply_connected_conformal_mapping",
|
| 71 |
"math.nnls_modulus_deblur",
|
|
|
|
| 72 |
"math.sigma_linear_diophantine",
|
| 73 |
"math.sparse_hyperbolic_recovery_2d",
|
| 74 |
"math.sparse_nonlinear_interaction_recovery",
|
| 75 |
"math.symplectic_integrators",
|
| 76 |
+
"math.ucb_q_learning_regret",
|
| 77 |
+
"physics.a2_trinion_full_index",
|
| 78 |
"physics.ag_strip_dielectric_inference",
|
| 79 |
+
"physics.broadband_maxwell_inverse_design",
|
| 80 |
"physics.dp_contact_process",
|
| 81 |
"physics.ks_2d_anisotropic",
|
| 82 |
"physics.kt_phase_transition_2d_xy_model",
|
|
|
|
| 84 |
"physics.lid_driven_cavity_flow",
|
| 85 |
"physics.paired_nanorod_negative_index_retrieval",
|
| 86 |
"physics.vpfm_leapfrog",
|
| 87 |
+
"robotics.2D_robot_landmark_EKF_SLAM",
|
| 88 |
"robotics.lidar_imu_time_offset_yaw_calibration",
|
| 89 |
"robotics.multiregime_icp_degeneracy",
|
| 90 |
"robotics.occupancy_grid_mapping_correlated_clutter",
|
|
|
|
| 95 |
],
|
| 96 |
"generated_at_utc": "2026-08-03T08:14:22Z",
|
| 97 |
"license": "Apache-2.0",
|
| 98 |
+
"license_file": "LICENSE",
|
| 99 |
+
"additional_task_sources": {
|
| 100 |
+
"accepted_v2_20261110": {
|
| 101 |
+
"task_ids": [
|
| 102 |
+
"biology.mixed_perturbation_gene_transfer",
|
| 103 |
+
"biology.subtle_regulatory_motif_discovery",
|
| 104 |
+
"biostat.adaptive_cohort_transport",
|
| 105 |
+
"computer_science.generate_conforming_hexahedral_elements_with_planar_faces",
|
| 106 |
+
"computer_science.planted_clique_community_recovery",
|
| 107 |
+
"computer_science.retry_storm_fault_localization",
|
| 108 |
+
"earth_science.flood_peak_reconstruction",
|
| 109 |
+
"earth_science.hysteretic_snowmelt_runoff_forecast",
|
| 110 |
+
"earth_science.robust_model_inversion_of_insar_deformation_at_an_active_volcano",
|
| 111 |
+
"electrical_engineering.battery_transient_thermal_forecast",
|
| 112 |
+
"electrical_engineering.vector_pair_limits",
|
| 113 |
+
"finance.robust_arbitrage_free_option_surface",
|
| 114 |
+
"materials.fcc_tilt_gb_family",
|
| 115 |
+
"physics.broadband_maxwell_inverse_design",
|
| 116 |
+
"robotics.2D_robot_landmark_EKF_SLAM"
|
| 117 |
+
],
|
| 118 |
+
"source_commit": "12f3d87d99848d0b0b3539a8777eed8448d1593e",
|
| 119 |
+
"seed": 31415,
|
| 120 |
+
"public_scorer_pr": "https://github.com/apexin-ai/ASI-Bench/pull/36",
|
| 121 |
+
"private_task_pr": "https://github.com/apexin-ai/Agent-AI4Sci-Bench/pull/119"
|
| 122 |
+
},
|
| 123 |
+
"accepted_v2_20261117": {
|
| 124 |
+
"task_ids": [
|
| 125 |
+
"computer_science.qp_sensitivity_audit",
|
| 126 |
+
"computer_science.minimal_latent_realization",
|
| 127 |
+
"computer_science.heterogeneous_click_log_counterfactuals",
|
| 128 |
+
"computer_science.signed_ising_log_partition",
|
| 129 |
+
"physics.a2_trinion_full_index",
|
| 130 |
+
"biology.unseen_perturbation_rna_response"
|
| 131 |
+
],
|
| 132 |
+
"seed": 31415,
|
| 133 |
+
"source_commit": "e688d3b1ec57632b66415cc0db564bee526064f2",
|
| 134 |
+
"private_task_pr": "https://github.com/apexin-ai/Agent-AI4Sci-Bench/pull/120",
|
| 135 |
+
"public_scorer_pr": "https://github.com/apexin-ai/ASI-Bench/pull/37",
|
| 136 |
+
"source_status": "Six accepted tasks in GitHub PRs; main merge pending user approval",
|
| 137 |
+
"fixed_instance_notes": {
|
| 138 |
+
"biology.unseen_perturbation_rna_response": "Original fixed measured split; harness seed31415 does not randomize the scientific split."
|
| 139 |
+
}
|
| 140 |
+
}
|
| 141 |
+
},
|
| 142 |
+
"updated_at_utc": "2026-11-17T10:50:01.436015+00:00"
|
| 143 |
}
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/data/experiments.json
ADDED
|
@@ -0,0 +1,1054 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"calibration_interventions": [
|
| 3 |
+
{
|
| 4 |
+
"id": "cal_00",
|
| 5 |
+
"rna_clamps": [
|
| 6 |
+
{
|
| 7 |
+
"gene": 8,
|
| 8 |
+
"value": 0.0
|
| 9 |
+
},
|
| 10 |
+
{
|
| 11 |
+
"gene": 3,
|
| 12 |
+
"value": 3.0
|
| 13 |
+
}
|
| 14 |
+
],
|
| 15 |
+
"inactive_regulators": [
|
| 16 |
+
14
|
| 17 |
+
]
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
"id": "cal_01",
|
| 21 |
+
"rna_clamps": [
|
| 22 |
+
{
|
| 23 |
+
"gene": 8,
|
| 24 |
+
"value": 3.0
|
| 25 |
+
},
|
| 26 |
+
{
|
| 27 |
+
"gene": 7,
|
| 28 |
+
"value": 0.0
|
| 29 |
+
}
|
| 30 |
+
],
|
| 31 |
+
"inactive_regulators": [
|
| 32 |
+
0
|
| 33 |
+
]
|
| 34 |
+
},
|
| 35 |
+
{
|
| 36 |
+
"id": "cal_02",
|
| 37 |
+
"rna_clamps": [
|
| 38 |
+
{
|
| 39 |
+
"gene": 24,
|
| 40 |
+
"value": 0.0
|
| 41 |
+
},
|
| 42 |
+
{
|
| 43 |
+
"gene": 4,
|
| 44 |
+
"value": 3.0
|
| 45 |
+
}
|
| 46 |
+
],
|
| 47 |
+
"inactive_regulators": [
|
| 48 |
+
2
|
| 49 |
+
]
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"id": "cal_03",
|
| 53 |
+
"rna_clamps": [
|
| 54 |
+
{
|
| 55 |
+
"gene": 1,
|
| 56 |
+
"value": 0.0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"gene": 24,
|
| 60 |
+
"value": 3.0
|
| 61 |
+
}
|
| 62 |
+
],
|
| 63 |
+
"inactive_regulators": [
|
| 64 |
+
19
|
| 65 |
+
]
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"id": "cal_04",
|
| 69 |
+
"rna_clamps": [
|
| 70 |
+
{
|
| 71 |
+
"gene": 19,
|
| 72 |
+
"value": 3.0
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"gene": 21,
|
| 76 |
+
"value": 0.0
|
| 77 |
+
}
|
| 78 |
+
],
|
| 79 |
+
"inactive_regulators": [
|
| 80 |
+
24
|
| 81 |
+
]
|
| 82 |
+
},
|
| 83 |
+
{
|
| 84 |
+
"id": "cal_05",
|
| 85 |
+
"rna_clamps": [
|
| 86 |
+
{
|
| 87 |
+
"gene": 20,
|
| 88 |
+
"value": 0.0
|
| 89 |
+
},
|
| 90 |
+
{
|
| 91 |
+
"gene": 24,
|
| 92 |
+
"value": 3.0
|
| 93 |
+
}
|
| 94 |
+
],
|
| 95 |
+
"inactive_regulators": [
|
| 96 |
+
5
|
| 97 |
+
]
|
| 98 |
+
},
|
| 99 |
+
{
|
| 100 |
+
"id": "cal_06",
|
| 101 |
+
"rna_clamps": [
|
| 102 |
+
{
|
| 103 |
+
"gene": 0,
|
| 104 |
+
"value": 0.0
|
| 105 |
+
},
|
| 106 |
+
{
|
| 107 |
+
"gene": 11,
|
| 108 |
+
"value": 3.0
|
| 109 |
+
}
|
| 110 |
+
],
|
| 111 |
+
"inactive_regulators": [
|
| 112 |
+
22
|
| 113 |
+
]
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"id": "cal_07",
|
| 117 |
+
"rna_clamps": [
|
| 118 |
+
{
|
| 119 |
+
"gene": 22,
|
| 120 |
+
"value": 3.0
|
| 121 |
+
},
|
| 122 |
+
{
|
| 123 |
+
"gene": 21,
|
| 124 |
+
"value": 0.0
|
| 125 |
+
}
|
| 126 |
+
],
|
| 127 |
+
"inactive_regulators": [
|
| 128 |
+
1
|
| 129 |
+
]
|
| 130 |
+
},
|
| 131 |
+
{
|
| 132 |
+
"id": "cal_08",
|
| 133 |
+
"rna_clamps": [
|
| 134 |
+
{
|
| 135 |
+
"gene": 3,
|
| 136 |
+
"value": 0.0
|
| 137 |
+
},
|
| 138 |
+
{
|
| 139 |
+
"gene": 18,
|
| 140 |
+
"value": 3.0
|
| 141 |
+
}
|
| 142 |
+
],
|
| 143 |
+
"inactive_regulators": [
|
| 144 |
+
6
|
| 145 |
+
]
|
| 146 |
+
},
|
| 147 |
+
{
|
| 148 |
+
"id": "cal_09",
|
| 149 |
+
"rna_clamps": [
|
| 150 |
+
{
|
| 151 |
+
"gene": 10,
|
| 152 |
+
"value": 0.0
|
| 153 |
+
},
|
| 154 |
+
{
|
| 155 |
+
"gene": 12,
|
| 156 |
+
"value": 3.0
|
| 157 |
+
}
|
| 158 |
+
],
|
| 159 |
+
"inactive_regulators": [
|
| 160 |
+
11
|
| 161 |
+
]
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"id": "cal_10",
|
| 165 |
+
"rna_clamps": [
|
| 166 |
+
{
|
| 167 |
+
"gene": 7,
|
| 168 |
+
"value": 3.0
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"gene": 3,
|
| 172 |
+
"value": 0.0
|
| 173 |
+
}
|
| 174 |
+
],
|
| 175 |
+
"inactive_regulators": [
|
| 176 |
+
23
|
| 177 |
+
]
|
| 178 |
+
},
|
| 179 |
+
{
|
| 180 |
+
"id": "cal_11",
|
| 181 |
+
"rna_clamps": [
|
| 182 |
+
{
|
| 183 |
+
"gene": 1,
|
| 184 |
+
"value": 0.0
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"gene": 17,
|
| 188 |
+
"value": 3.0
|
| 189 |
+
}
|
| 190 |
+
],
|
| 191 |
+
"inactive_regulators": [
|
| 192 |
+
2
|
| 193 |
+
]
|
| 194 |
+
},
|
| 195 |
+
{
|
| 196 |
+
"id": "cal_12",
|
| 197 |
+
"rna_clamps": [
|
| 198 |
+
{
|
| 199 |
+
"gene": 6,
|
| 200 |
+
"value": 0.0
|
| 201 |
+
},
|
| 202 |
+
{
|
| 203 |
+
"gene": 24,
|
| 204 |
+
"value": 3.0
|
| 205 |
+
}
|
| 206 |
+
],
|
| 207 |
+
"inactive_regulators": [
|
| 208 |
+
2
|
| 209 |
+
]
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"id": "cal_13",
|
| 213 |
+
"rna_clamps": [
|
| 214 |
+
{
|
| 215 |
+
"gene": 14,
|
| 216 |
+
"value": 3.0
|
| 217 |
+
},
|
| 218 |
+
{
|
| 219 |
+
"gene": 6,
|
| 220 |
+
"value": 0.0
|
| 221 |
+
}
|
| 222 |
+
],
|
| 223 |
+
"inactive_regulators": [
|
| 224 |
+
23
|
| 225 |
+
]
|
| 226 |
+
},
|
| 227 |
+
{
|
| 228 |
+
"id": "cal_14",
|
| 229 |
+
"rna_clamps": [
|
| 230 |
+
{
|
| 231 |
+
"gene": 19,
|
| 232 |
+
"value": 0.0
|
| 233 |
+
},
|
| 234 |
+
{
|
| 235 |
+
"gene": 22,
|
| 236 |
+
"value": 3.0
|
| 237 |
+
}
|
| 238 |
+
],
|
| 239 |
+
"inactive_regulators": [
|
| 240 |
+
13
|
| 241 |
+
]
|
| 242 |
+
}
|
| 243 |
+
],
|
| 244 |
+
"query_protocols": [
|
| 245 |
+
{
|
| 246 |
+
"id": "query_00",
|
| 247 |
+
"segments": [
|
| 248 |
+
{
|
| 249 |
+
"duration": 1.5,
|
| 250 |
+
"rna_clamps": [
|
| 251 |
+
{
|
| 252 |
+
"gene": 8,
|
| 253 |
+
"value": 0.0
|
| 254 |
+
},
|
| 255 |
+
{
|
| 256 |
+
"gene": 14,
|
| 257 |
+
"value": 2.7043698608490585
|
| 258 |
+
}
|
| 259 |
+
],
|
| 260 |
+
"inactive_regulators": [
|
| 261 |
+
16
|
| 262 |
+
]
|
| 263 |
+
},
|
| 264 |
+
{
|
| 265 |
+
"duration": 1.5,
|
| 266 |
+
"rna_clamps": [
|
| 267 |
+
{
|
| 268 |
+
"gene": 12,
|
| 269 |
+
"value": 2.763962066472951
|
| 270 |
+
}
|
| 271 |
+
],
|
| 272 |
+
"inactive_regulators": [
|
| 273 |
+
6
|
| 274 |
+
]
|
| 275 |
+
}
|
| 276 |
+
]
|
| 277 |
+
},
|
| 278 |
+
{
|
| 279 |
+
"id": "query_01",
|
| 280 |
+
"segments": [
|
| 281 |
+
{
|
| 282 |
+
"duration": 1.5,
|
| 283 |
+
"rna_clamps": [
|
| 284 |
+
{
|
| 285 |
+
"gene": 18,
|
| 286 |
+
"value": 2.5061326683603435
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"gene": 0,
|
| 290 |
+
"value": 0.0
|
| 291 |
+
}
|
| 292 |
+
],
|
| 293 |
+
"inactive_regulators": [
|
| 294 |
+
6
|
| 295 |
+
]
|
| 296 |
+
},
|
| 297 |
+
{
|
| 298 |
+
"duration": 1.5,
|
| 299 |
+
"rna_clamps": [
|
| 300 |
+
{
|
| 301 |
+
"gene": 18,
|
| 302 |
+
"value": 0.0
|
| 303 |
+
}
|
| 304 |
+
],
|
| 305 |
+
"inactive_regulators": [
|
| 306 |
+
24
|
| 307 |
+
]
|
| 308 |
+
}
|
| 309 |
+
]
|
| 310 |
+
},
|
| 311 |
+
{
|
| 312 |
+
"id": "query_02",
|
| 313 |
+
"segments": [
|
| 314 |
+
{
|
| 315 |
+
"duration": 1.5,
|
| 316 |
+
"rna_clamps": [
|
| 317 |
+
{
|
| 318 |
+
"gene": 15,
|
| 319 |
+
"value": 0.0
|
| 320 |
+
},
|
| 321 |
+
{
|
| 322 |
+
"gene": 7,
|
| 323 |
+
"value": 3.4965527829644927
|
| 324 |
+
}
|
| 325 |
+
],
|
| 326 |
+
"inactive_regulators": [
|
| 327 |
+
13
|
| 328 |
+
]
|
| 329 |
+
},
|
| 330 |
+
{
|
| 331 |
+
"duration": 1.5,
|
| 332 |
+
"rna_clamps": [
|
| 333 |
+
{
|
| 334 |
+
"gene": 19,
|
| 335 |
+
"value": 0.0
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"gene": 23,
|
| 339 |
+
"value": 3.8143059697761803
|
| 340 |
+
}
|
| 341 |
+
],
|
| 342 |
+
"inactive_regulators": []
|
| 343 |
+
}
|
| 344 |
+
]
|
| 345 |
+
},
|
| 346 |
+
{
|
| 347 |
+
"id": "query_03",
|
| 348 |
+
"segments": [
|
| 349 |
+
{
|
| 350 |
+
"duration": 1.5,
|
| 351 |
+
"rna_clamps": [
|
| 352 |
+
{
|
| 353 |
+
"gene": 6,
|
| 354 |
+
"value": 0.0
|
| 355 |
+
},
|
| 356 |
+
{
|
| 357 |
+
"gene": 22,
|
| 358 |
+
"value": 2.951148648158107
|
| 359 |
+
}
|
| 360 |
+
],
|
| 361 |
+
"inactive_regulators": [
|
| 362 |
+
18
|
| 363 |
+
]
|
| 364 |
+
},
|
| 365 |
+
{
|
| 366 |
+
"duration": 1.5,
|
| 367 |
+
"rna_clamps": [
|
| 368 |
+
{
|
| 369 |
+
"gene": 6,
|
| 370 |
+
"value": 3.0195089206626937
|
| 371 |
+
}
|
| 372 |
+
],
|
| 373 |
+
"inactive_regulators": [
|
| 374 |
+
4
|
| 375 |
+
]
|
| 376 |
+
}
|
| 377 |
+
]
|
| 378 |
+
},
|
| 379 |
+
{
|
| 380 |
+
"id": "query_04",
|
| 381 |
+
"segments": [
|
| 382 |
+
{
|
| 383 |
+
"duration": 1.5,
|
| 384 |
+
"rna_clamps": [
|
| 385 |
+
{
|
| 386 |
+
"gene": 5,
|
| 387 |
+
"value": 3.2183160732839506
|
| 388 |
+
},
|
| 389 |
+
{
|
| 390 |
+
"gene": 10,
|
| 391 |
+
"value": 0.0
|
| 392 |
+
}
|
| 393 |
+
],
|
| 394 |
+
"inactive_regulators": [
|
| 395 |
+
7
|
| 396 |
+
]
|
| 397 |
+
},
|
| 398 |
+
{
|
| 399 |
+
"duration": 1.5,
|
| 400 |
+
"rna_clamps": [
|
| 401 |
+
{
|
| 402 |
+
"gene": 16,
|
| 403 |
+
"value": 0.0
|
| 404 |
+
}
|
| 405 |
+
],
|
| 406 |
+
"inactive_regulators": [
|
| 407 |
+
3
|
| 408 |
+
]
|
| 409 |
+
}
|
| 410 |
+
]
|
| 411 |
+
},
|
| 412 |
+
{
|
| 413 |
+
"id": "query_05",
|
| 414 |
+
"segments": [
|
| 415 |
+
{
|
| 416 |
+
"duration": 1.5,
|
| 417 |
+
"rna_clamps": [
|
| 418 |
+
{
|
| 419 |
+
"gene": 10,
|
| 420 |
+
"value": 0.0
|
| 421 |
+
},
|
| 422 |
+
{
|
| 423 |
+
"gene": 8,
|
| 424 |
+
"value": 3.366999612287615
|
| 425 |
+
}
|
| 426 |
+
],
|
| 427 |
+
"inactive_regulators": [
|
| 428 |
+
7
|
| 429 |
+
]
|
| 430 |
+
},
|
| 431 |
+
{
|
| 432 |
+
"duration": 1.5,
|
| 433 |
+
"rna_clamps": [
|
| 434 |
+
{
|
| 435 |
+
"gene": 13,
|
| 436 |
+
"value": 0.0
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"gene": 9,
|
| 440 |
+
"value": 3.800035901186815
|
| 441 |
+
}
|
| 442 |
+
],
|
| 443 |
+
"inactive_regulators": []
|
| 444 |
+
}
|
| 445 |
+
]
|
| 446 |
+
},
|
| 447 |
+
{
|
| 448 |
+
"id": "query_06",
|
| 449 |
+
"segments": [
|
| 450 |
+
{
|
| 451 |
+
"duration": 1.5,
|
| 452 |
+
"rna_clamps": [
|
| 453 |
+
{
|
| 454 |
+
"gene": 1,
|
| 455 |
+
"value": 0.0
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"gene": 3,
|
| 459 |
+
"value": 3.3847784912252177
|
| 460 |
+
}
|
| 461 |
+
],
|
| 462 |
+
"inactive_regulators": [
|
| 463 |
+
15
|
| 464 |
+
]
|
| 465 |
+
},
|
| 466 |
+
{
|
| 467 |
+
"duration": 1.5,
|
| 468 |
+
"rna_clamps": [
|
| 469 |
+
{
|
| 470 |
+
"gene": 23,
|
| 471 |
+
"value": 3.7169006408852323
|
| 472 |
+
}
|
| 473 |
+
],
|
| 474 |
+
"inactive_regulators": [
|
| 475 |
+
1
|
| 476 |
+
]
|
| 477 |
+
}
|
| 478 |
+
]
|
| 479 |
+
},
|
| 480 |
+
{
|
| 481 |
+
"id": "query_07",
|
| 482 |
+
"segments": [
|
| 483 |
+
{
|
| 484 |
+
"duration": 1.5,
|
| 485 |
+
"rna_clamps": [
|
| 486 |
+
{
|
| 487 |
+
"gene": 21,
|
| 488 |
+
"value": 3.676363639129728
|
| 489 |
+
},
|
| 490 |
+
{
|
| 491 |
+
"gene": 20,
|
| 492 |
+
"value": 0.0
|
| 493 |
+
}
|
| 494 |
+
],
|
| 495 |
+
"inactive_regulators": [
|
| 496 |
+
14
|
| 497 |
+
]
|
| 498 |
+
},
|
| 499 |
+
{
|
| 500 |
+
"duration": 1.5,
|
| 501 |
+
"rna_clamps": [
|
| 502 |
+
{
|
| 503 |
+
"gene": 9,
|
| 504 |
+
"value": 0.0
|
| 505 |
+
}
|
| 506 |
+
],
|
| 507 |
+
"inactive_regulators": [
|
| 508 |
+
3
|
| 509 |
+
]
|
| 510 |
+
}
|
| 511 |
+
]
|
| 512 |
+
},
|
| 513 |
+
{
|
| 514 |
+
"id": "query_08",
|
| 515 |
+
"segments": [
|
| 516 |
+
{
|
| 517 |
+
"duration": 1.5,
|
| 518 |
+
"rna_clamps": [
|
| 519 |
+
{
|
| 520 |
+
"gene": 6,
|
| 521 |
+
"value": 0.0
|
| 522 |
+
},
|
| 523 |
+
{
|
| 524 |
+
"gene": 0,
|
| 525 |
+
"value": 3.105663000006595
|
| 526 |
+
}
|
| 527 |
+
],
|
| 528 |
+
"inactive_regulators": [
|
| 529 |
+
15
|
| 530 |
+
]
|
| 531 |
+
},
|
| 532 |
+
{
|
| 533 |
+
"duration": 1.5,
|
| 534 |
+
"rna_clamps": [
|
| 535 |
+
{
|
| 536 |
+
"gene": 11,
|
| 537 |
+
"value": 0.0
|
| 538 |
+
},
|
| 539 |
+
{
|
| 540 |
+
"gene": 3,
|
| 541 |
+
"value": 3.4366431350711633
|
| 542 |
+
}
|
| 543 |
+
],
|
| 544 |
+
"inactive_regulators": []
|
| 545 |
+
}
|
| 546 |
+
]
|
| 547 |
+
},
|
| 548 |
+
{
|
| 549 |
+
"id": "query_09",
|
| 550 |
+
"segments": [
|
| 551 |
+
{
|
| 552 |
+
"duration": 1.5,
|
| 553 |
+
"rna_clamps": [
|
| 554 |
+
{
|
| 555 |
+
"gene": 10,
|
| 556 |
+
"value": 0.0
|
| 557 |
+
},
|
| 558 |
+
{
|
| 559 |
+
"gene": 0,
|
| 560 |
+
"value": 3.2133156404353906
|
| 561 |
+
}
|
| 562 |
+
],
|
| 563 |
+
"inactive_regulators": [
|
| 564 |
+
15
|
| 565 |
+
]
|
| 566 |
+
},
|
| 567 |
+
{
|
| 568 |
+
"duration": 1.5,
|
| 569 |
+
"rna_clamps": [
|
| 570 |
+
{
|
| 571 |
+
"gene": 0,
|
| 572 |
+
"value": 3.8558909512515456
|
| 573 |
+
}
|
| 574 |
+
],
|
| 575 |
+
"inactive_regulators": [
|
| 576 |
+
23
|
| 577 |
+
]
|
| 578 |
+
}
|
| 579 |
+
]
|
| 580 |
+
},
|
| 581 |
+
{
|
| 582 |
+
"id": "query_10",
|
| 583 |
+
"segments": [
|
| 584 |
+
{
|
| 585 |
+
"duration": 1.5,
|
| 586 |
+
"rna_clamps": [
|
| 587 |
+
{
|
| 588 |
+
"gene": 1,
|
| 589 |
+
"value": 3.5155744565634404
|
| 590 |
+
},
|
| 591 |
+
{
|
| 592 |
+
"gene": 17,
|
| 593 |
+
"value": 0.0
|
| 594 |
+
}
|
| 595 |
+
],
|
| 596 |
+
"inactive_regulators": [
|
| 597 |
+
15
|
| 598 |
+
]
|
| 599 |
+
},
|
| 600 |
+
{
|
| 601 |
+
"duration": 1.5,
|
| 602 |
+
"rna_clamps": [
|
| 603 |
+
{
|
| 604 |
+
"gene": 11,
|
| 605 |
+
"value": 0.0
|
| 606 |
+
}
|
| 607 |
+
],
|
| 608 |
+
"inactive_regulators": [
|
| 609 |
+
4
|
| 610 |
+
]
|
| 611 |
+
}
|
| 612 |
+
]
|
| 613 |
+
},
|
| 614 |
+
{
|
| 615 |
+
"id": "query_11",
|
| 616 |
+
"segments": [
|
| 617 |
+
{
|
| 618 |
+
"duration": 1.5,
|
| 619 |
+
"rna_clamps": [
|
| 620 |
+
{
|
| 621 |
+
"gene": 19,
|
| 622 |
+
"value": 0.0
|
| 623 |
+
},
|
| 624 |
+
{
|
| 625 |
+
"gene": 21,
|
| 626 |
+
"value": 2.800116047684265
|
| 627 |
+
}
|
| 628 |
+
],
|
| 629 |
+
"inactive_regulators": [
|
| 630 |
+
23
|
| 631 |
+
]
|
| 632 |
+
},
|
| 633 |
+
{
|
| 634 |
+
"duration": 1.5,
|
| 635 |
+
"rna_clamps": [
|
| 636 |
+
{
|
| 637 |
+
"gene": 3,
|
| 638 |
+
"value": 0.0
|
| 639 |
+
},
|
| 640 |
+
{
|
| 641 |
+
"gene": 22,
|
| 642 |
+
"value": 2.7721718752048585
|
| 643 |
+
}
|
| 644 |
+
],
|
| 645 |
+
"inactive_regulators": []
|
| 646 |
+
}
|
| 647 |
+
]
|
| 648 |
+
},
|
| 649 |
+
{
|
| 650 |
+
"id": "query_12",
|
| 651 |
+
"segments": [
|
| 652 |
+
{
|
| 653 |
+
"duration": 1.5,
|
| 654 |
+
"rna_clamps": [
|
| 655 |
+
{
|
| 656 |
+
"gene": 18,
|
| 657 |
+
"value": 0.0
|
| 658 |
+
},
|
| 659 |
+
{
|
| 660 |
+
"gene": 3,
|
| 661 |
+
"value": 2.723249363956767
|
| 662 |
+
}
|
| 663 |
+
],
|
| 664 |
+
"inactive_regulators": [
|
| 665 |
+
22
|
| 666 |
+
]
|
| 667 |
+
},
|
| 668 |
+
{
|
| 669 |
+
"duration": 1.5,
|
| 670 |
+
"rna_clamps": [
|
| 671 |
+
{
|
| 672 |
+
"gene": 23,
|
| 673 |
+
"value": 3.162410848385857
|
| 674 |
+
}
|
| 675 |
+
],
|
| 676 |
+
"inactive_regulators": [
|
| 677 |
+
4
|
| 678 |
+
]
|
| 679 |
+
}
|
| 680 |
+
]
|
| 681 |
+
},
|
| 682 |
+
{
|
| 683 |
+
"id": "query_13",
|
| 684 |
+
"segments": [
|
| 685 |
+
{
|
| 686 |
+
"duration": 1.5,
|
| 687 |
+
"rna_clamps": [
|
| 688 |
+
{
|
| 689 |
+
"gene": 13,
|
| 690 |
+
"value": 3.6085470443833634
|
| 691 |
+
},
|
| 692 |
+
{
|
| 693 |
+
"gene": 3,
|
| 694 |
+
"value": 0.0
|
| 695 |
+
}
|
| 696 |
+
],
|
| 697 |
+
"inactive_regulators": [
|
| 698 |
+
20
|
| 699 |
+
]
|
| 700 |
+
},
|
| 701 |
+
{
|
| 702 |
+
"duration": 1.5,
|
| 703 |
+
"rna_clamps": [
|
| 704 |
+
{
|
| 705 |
+
"gene": 24,
|
| 706 |
+
"value": 0.0
|
| 707 |
+
}
|
| 708 |
+
],
|
| 709 |
+
"inactive_regulators": [
|
| 710 |
+
21
|
| 711 |
+
]
|
| 712 |
+
}
|
| 713 |
+
]
|
| 714 |
+
},
|
| 715 |
+
{
|
| 716 |
+
"id": "query_14",
|
| 717 |
+
"segments": [
|
| 718 |
+
{
|
| 719 |
+
"duration": 1.5,
|
| 720 |
+
"rna_clamps": [
|
| 721 |
+
{
|
| 722 |
+
"gene": 8,
|
| 723 |
+
"value": 0.0
|
| 724 |
+
},
|
| 725 |
+
{
|
| 726 |
+
"gene": 10,
|
| 727 |
+
"value": 3.7696297134561774
|
| 728 |
+
}
|
| 729 |
+
],
|
| 730 |
+
"inactive_regulators": [
|
| 731 |
+
6
|
| 732 |
+
]
|
| 733 |
+
},
|
| 734 |
+
{
|
| 735 |
+
"duration": 1.5,
|
| 736 |
+
"rna_clamps": [
|
| 737 |
+
{
|
| 738 |
+
"gene": 17,
|
| 739 |
+
"value": 0.0
|
| 740 |
+
},
|
| 741 |
+
{
|
| 742 |
+
"gene": 22,
|
| 743 |
+
"value": 2.9509735502140155
|
| 744 |
+
}
|
| 745 |
+
],
|
| 746 |
+
"inactive_regulators": []
|
| 747 |
+
}
|
| 748 |
+
]
|
| 749 |
+
},
|
| 750 |
+
{
|
| 751 |
+
"id": "query_15",
|
| 752 |
+
"segments": [
|
| 753 |
+
{
|
| 754 |
+
"duration": 1.5,
|
| 755 |
+
"rna_clamps": [
|
| 756 |
+
{
|
| 757 |
+
"gene": 14,
|
| 758 |
+
"value": 0.0
|
| 759 |
+
},
|
| 760 |
+
{
|
| 761 |
+
"gene": 2,
|
| 762 |
+
"value": 2.698235314133263
|
| 763 |
+
}
|
| 764 |
+
],
|
| 765 |
+
"inactive_regulators": [
|
| 766 |
+
5
|
| 767 |
+
]
|
| 768 |
+
},
|
| 769 |
+
{
|
| 770 |
+
"duration": 1.5,
|
| 771 |
+
"rna_clamps": [
|
| 772 |
+
{
|
| 773 |
+
"gene": 4,
|
| 774 |
+
"value": 3.4999719450697633
|
| 775 |
+
}
|
| 776 |
+
],
|
| 777 |
+
"inactive_regulators": [
|
| 778 |
+
1
|
| 779 |
+
]
|
| 780 |
+
}
|
| 781 |
+
]
|
| 782 |
+
},
|
| 783 |
+
{
|
| 784 |
+
"id": "query_16",
|
| 785 |
+
"segments": [
|
| 786 |
+
{
|
| 787 |
+
"duration": 1.5,
|
| 788 |
+
"rna_clamps": [
|
| 789 |
+
{
|
| 790 |
+
"gene": 17,
|
| 791 |
+
"value": 3.7204939067385463
|
| 792 |
+
},
|
| 793 |
+
{
|
| 794 |
+
"gene": 13,
|
| 795 |
+
"value": 0.0
|
| 796 |
+
}
|
| 797 |
+
],
|
| 798 |
+
"inactive_regulators": [
|
| 799 |
+
0
|
| 800 |
+
]
|
| 801 |
+
},
|
| 802 |
+
{
|
| 803 |
+
"duration": 1.5,
|
| 804 |
+
"rna_clamps": [
|
| 805 |
+
{
|
| 806 |
+
"gene": 18,
|
| 807 |
+
"value": 0.0
|
| 808 |
+
}
|
| 809 |
+
],
|
| 810 |
+
"inactive_regulators": [
|
| 811 |
+
11
|
| 812 |
+
]
|
| 813 |
+
}
|
| 814 |
+
]
|
| 815 |
+
},
|
| 816 |
+
{
|
| 817 |
+
"id": "query_17",
|
| 818 |
+
"segments": [
|
| 819 |
+
{
|
| 820 |
+
"duration": 1.5,
|
| 821 |
+
"rna_clamps": [
|
| 822 |
+
{
|
| 823 |
+
"gene": 4,
|
| 824 |
+
"value": 0.0
|
| 825 |
+
},
|
| 826 |
+
{
|
| 827 |
+
"gene": 0,
|
| 828 |
+
"value": 3.2381149023850897
|
| 829 |
+
}
|
| 830 |
+
],
|
| 831 |
+
"inactive_regulators": [
|
| 832 |
+
1
|
| 833 |
+
]
|
| 834 |
+
},
|
| 835 |
+
{
|
| 836 |
+
"duration": 1.5,
|
| 837 |
+
"rna_clamps": [
|
| 838 |
+
{
|
| 839 |
+
"gene": 8,
|
| 840 |
+
"value": 0.0
|
| 841 |
+
},
|
| 842 |
+
{
|
| 843 |
+
"gene": 1,
|
| 844 |
+
"value": 3.7843402735863787
|
| 845 |
+
}
|
| 846 |
+
],
|
| 847 |
+
"inactive_regulators": []
|
| 848 |
+
}
|
| 849 |
+
]
|
| 850 |
+
},
|
| 851 |
+
{
|
| 852 |
+
"id": "query_18",
|
| 853 |
+
"segments": [
|
| 854 |
+
{
|
| 855 |
+
"duration": 1.5,
|
| 856 |
+
"rna_clamps": [
|
| 857 |
+
{
|
| 858 |
+
"gene": 17,
|
| 859 |
+
"value": 0.0
|
| 860 |
+
},
|
| 861 |
+
{
|
| 862 |
+
"gene": 9,
|
| 863 |
+
"value": 3.5450690106494367
|
| 864 |
+
}
|
| 865 |
+
],
|
| 866 |
+
"inactive_regulators": [
|
| 867 |
+
19
|
| 868 |
+
]
|
| 869 |
+
},
|
| 870 |
+
{
|
| 871 |
+
"duration": 1.5,
|
| 872 |
+
"rna_clamps": [
|
| 873 |
+
{
|
| 874 |
+
"gene": 21,
|
| 875 |
+
"value": 3.870652602780213
|
| 876 |
+
}
|
| 877 |
+
],
|
| 878 |
+
"inactive_regulators": [
|
| 879 |
+
18
|
| 880 |
+
]
|
| 881 |
+
}
|
| 882 |
+
]
|
| 883 |
+
},
|
| 884 |
+
{
|
| 885 |
+
"id": "query_19",
|
| 886 |
+
"segments": [
|
| 887 |
+
{
|
| 888 |
+
"duration": 1.5,
|
| 889 |
+
"rna_clamps": [
|
| 890 |
+
{
|
| 891 |
+
"gene": 11,
|
| 892 |
+
"value": 2.596613925268655
|
| 893 |
+
},
|
| 894 |
+
{
|
| 895 |
+
"gene": 6,
|
| 896 |
+
"value": 0.0
|
| 897 |
+
}
|
| 898 |
+
],
|
| 899 |
+
"inactive_regulators": [
|
| 900 |
+
10
|
| 901 |
+
]
|
| 902 |
+
},
|
| 903 |
+
{
|
| 904 |
+
"duration": 1.5,
|
| 905 |
+
"rna_clamps": [
|
| 906 |
+
{
|
| 907 |
+
"gene": 4,
|
| 908 |
+
"value": 0.0
|
| 909 |
+
}
|
| 910 |
+
],
|
| 911 |
+
"inactive_regulators": [
|
| 912 |
+
8
|
| 913 |
+
]
|
| 914 |
+
}
|
| 915 |
+
]
|
| 916 |
+
},
|
| 917 |
+
{
|
| 918 |
+
"id": "query_20",
|
| 919 |
+
"segments": [
|
| 920 |
+
{
|
| 921 |
+
"duration": 1.5,
|
| 922 |
+
"rna_clamps": [
|
| 923 |
+
{
|
| 924 |
+
"gene": 15,
|
| 925 |
+
"value": 0.0
|
| 926 |
+
},
|
| 927 |
+
{
|
| 928 |
+
"gene": 20,
|
| 929 |
+
"value": 3.492846408340591
|
| 930 |
+
}
|
| 931 |
+
],
|
| 932 |
+
"inactive_regulators": [
|
| 933 |
+
12
|
| 934 |
+
]
|
| 935 |
+
},
|
| 936 |
+
{
|
| 937 |
+
"duration": 1.5,
|
| 938 |
+
"rna_clamps": [
|
| 939 |
+
{
|
| 940 |
+
"gene": 9,
|
| 941 |
+
"value": 0.0
|
| 942 |
+
},
|
| 943 |
+
{
|
| 944 |
+
"gene": 5,
|
| 945 |
+
"value": 3.76534802083711
|
| 946 |
+
}
|
| 947 |
+
],
|
| 948 |
+
"inactive_regulators": []
|
| 949 |
+
}
|
| 950 |
+
]
|
| 951 |
+
},
|
| 952 |
+
{
|
| 953 |
+
"id": "query_21",
|
| 954 |
+
"segments": [
|
| 955 |
+
{
|
| 956 |
+
"duration": 1.5,
|
| 957 |
+
"rna_clamps": [
|
| 958 |
+
{
|
| 959 |
+
"gene": 8,
|
| 960 |
+
"value": 0.0
|
| 961 |
+
},
|
| 962 |
+
{
|
| 963 |
+
"gene": 23,
|
| 964 |
+
"value": 3.8388356965261146
|
| 965 |
+
}
|
| 966 |
+
],
|
| 967 |
+
"inactive_regulators": [
|
| 968 |
+
15
|
| 969 |
+
]
|
| 970 |
+
},
|
| 971 |
+
{
|
| 972 |
+
"duration": 1.5,
|
| 973 |
+
"rna_clamps": [
|
| 974 |
+
{
|
| 975 |
+
"gene": 0,
|
| 976 |
+
"value": 2.9313493121933467
|
| 977 |
+
}
|
| 978 |
+
],
|
| 979 |
+
"inactive_regulators": [
|
| 980 |
+
7
|
| 981 |
+
]
|
| 982 |
+
}
|
| 983 |
+
]
|
| 984 |
+
},
|
| 985 |
+
{
|
| 986 |
+
"id": "query_22",
|
| 987 |
+
"segments": [
|
| 988 |
+
{
|
| 989 |
+
"duration": 1.5,
|
| 990 |
+
"rna_clamps": [
|
| 991 |
+
{
|
| 992 |
+
"gene": 4,
|
| 993 |
+
"value": 2.9883992269298836
|
| 994 |
+
},
|
| 995 |
+
{
|
| 996 |
+
"gene": 10,
|
| 997 |
+
"value": 0.0
|
| 998 |
+
}
|
| 999 |
+
],
|
| 1000 |
+
"inactive_regulators": [
|
| 1001 |
+
12
|
| 1002 |
+
]
|
| 1003 |
+
},
|
| 1004 |
+
{
|
| 1005 |
+
"duration": 1.5,
|
| 1006 |
+
"rna_clamps": [
|
| 1007 |
+
{
|
| 1008 |
+
"gene": 10,
|
| 1009 |
+
"value": 0.0
|
| 1010 |
+
}
|
| 1011 |
+
],
|
| 1012 |
+
"inactive_regulators": [
|
| 1013 |
+
1
|
| 1014 |
+
]
|
| 1015 |
+
}
|
| 1016 |
+
]
|
| 1017 |
+
},
|
| 1018 |
+
{
|
| 1019 |
+
"id": "query_23",
|
| 1020 |
+
"segments": [
|
| 1021 |
+
{
|
| 1022 |
+
"duration": 1.5,
|
| 1023 |
+
"rna_clamps": [
|
| 1024 |
+
{
|
| 1025 |
+
"gene": 23,
|
| 1026 |
+
"value": 0.0
|
| 1027 |
+
},
|
| 1028 |
+
{
|
| 1029 |
+
"gene": 10,
|
| 1030 |
+
"value": 2.837876642295724
|
| 1031 |
+
}
|
| 1032 |
+
],
|
| 1033 |
+
"inactive_regulators": [
|
| 1034 |
+
24
|
| 1035 |
+
]
|
| 1036 |
+
},
|
| 1037 |
+
{
|
| 1038 |
+
"duration": 1.5,
|
| 1039 |
+
"rna_clamps": [
|
| 1040 |
+
{
|
| 1041 |
+
"gene": 2,
|
| 1042 |
+
"value": 0.0
|
| 1043 |
+
},
|
| 1044 |
+
{
|
| 1045 |
+
"gene": 22,
|
| 1046 |
+
"value": 3.204412473175626
|
| 1047 |
+
}
|
| 1048 |
+
],
|
| 1049 |
+
"inactive_regulators": []
|
| 1050 |
+
}
|
| 1051 |
+
]
|
| 1052 |
+
}
|
| 1053 |
+
]
|
| 1054 |
+
}
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/data/observations.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ee9a0acf3e3f429ebf78791d10a2da0907184a997cc8eaca45d819c61e4a6064
|
| 3 |
+
size 139750
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/data/study.json
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema_version": 1,
|
| 3 |
+
"n_genes": 25,
|
| 4 |
+
"time_unit": "minute",
|
| 5 |
+
"rna_units": "Calibrated transcript abundance units; all 25 circuit transcripts are observed.",
|
| 6 |
+
"protein_units": "Calibrated functional-protein abundance units, on a common assay scale.",
|
| 7 |
+
"biology": "Closed synthetic transcription circuit. Transcript abundance and functional-protein abundance are distinct dynamical quantities. Transcripts compete for a finite shared translation resource. Direct promoter regulation is activating or repressing; no gene directly regulates its own promoter. Regulatory and resource laws are time homogeneous and shared across experiments, with fixed gene-specific basal synthesis and macromolecule turnover parameters. The following exact functional laws, including their constants, are known and apply at every prompt level. For RNA x_i and functional protein p_i, q_i(x)=12*x_i/(5+sum_j x_j), dp_i/dt=gamma_p_i*(q_i(x)-p_i), and H(p_i)=p_i^3/(0.65^3+p_i^3). W[i,j] uses target i and source j; positive means activation, negative means repression, and zero means no direct edge. Set f_ij=H(p_j) for an active activating parent, f_ij=1-H(p_j) for an active repressing parent, and f_ij=0 for absent or inactivated parents. RNA obeys dx_i/dt=b_i+sum_j abs(W[i,j])*f_ij+kappa_i*sum_{j<k} f_ij*f_ik-gamma_m_i*x_i, where the pair sum is over unordered pairs of distinct direct parents of target i. Inactivation removes a regulator from both the main and cooperative terms, while its own RNA/protein still evolve. RNA clamps override the corresponding RNA equation, immediately set that RNA at the boundary and hold it fixed; protein is never reset. The diagonal of W is zero and each target has at most three direct parents. The direct wiring, signs, weight magnitudes, positive basal rates b_i, nonnegative cooperative strengths kappa_i, and positive gene-specific decay rates gamma_m_i and gamma_p_i must be inferred from the public measurements. The constants 12, 5, 3, and 0.65 are fixed known quantities, not hidden inference targets. This is a controlled kinetic idealization, not observations from humans or animals.",
|
| 8 |
+
"measurements": "rna[r,t,i] records transcript i at times[t] in untreated time course r. initial_protein[r,i] is its separately measured protein abundance at time zero. rna has independent Gaussian measurement error with standard deviation observation_noise_relative*(0.2+true abundance), clipped below at zero. late_rna[0] and wildtype_protein give the noise-free untreated reference preparation. late_rna[1:] are transcript observations 45 minutes after the listed constant calibration interventions, with the same RNA noise law; these are late endpoints and need not be assumed equilibria.",
|
| 9 |
+
"decay_assays": "rna_retention[i,t] and protein_retention[i,t] are fractions remaining at decay_times[t] in independent first-order loss assays with new synthesis blocked. These fractions, initial protein values, and the reference preparation are exact calibration measurements. Fraction one corresponds to the start of the corresponding decay assay.",
|
| 10 |
+
"interventions": "rna_clamps instantaneously set the specified transcript abundances and maintain them throughout that segment. A clamp does not reset protein abundance. inactive_regulators removes the listed proteins contributions at every promoter, including all of their activating and repressing promoter actions, while their own transcript/protein dynamics continue. Each segment lists the complete current controls; omitted previous controls are released. At a boundary, states carry forward, except transcripts explicitly reset by new clamps. Releasing a clamp does not reset that transcript. Calibration experiments start independently from the reference preparation and maintain their controls for 45 minutes.",
|
| 11 |
+
"prediction": "Each query starts independently from late_rna[0] and wildtype_protein. Predict noise-free transcript abundances at six equally spaced positive times in each segment, duration/6 through duration. Concatenate segment samples in order; protein states continue across boundaries. The two segments give twelve observations per query.",
|
| 12 |
+
"direct_edge": "weights[target,source] represents direct action of source protein at the target promoter: positive activation, negative repression, zero absence. Effects transmitted through shared translation capacity or other genes are not direct edges. Hubs are the two largest direct out-degree genes, ties resolved by smaller gene index. Gene indices run from 0 through 24.",
|
| 13 |
+
"observation_noise_relative": 0.003
|
| 14 |
+
}
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/instance_meta.json
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"params_used": {
|
| 3 |
+
"seed": 31415
|
| 4 |
+
},
|
| 5 |
+
"input_files": [
|
| 6 |
+
"study.json",
|
| 7 |
+
"observations.npz",
|
| 8 |
+
"experiments.json"
|
| 9 |
+
],
|
| 10 |
+
"reference_files": [
|
| 11 |
+
"ground_truth.npz",
|
| 12 |
+
"predictions.npz",
|
| 13 |
+
"analysis.py",
|
| 14 |
+
"report.md"
|
| 15 |
+
],
|
| 16 |
+
"generation_time_seconds": 1.7986609255895019,
|
| 17 |
+
"data_origin": "Original controlled synthetic circuit; no human or animal data"
|
| 18 |
+
}
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/prompt_b1.md
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Recovering a Gene Network from Mixed Perturbations
|
| 2 |
+
|
| 3 |
+
## Scientific goal
|
| 4 |
+
|
| 5 |
+
Infer the signed **direct** transcriptional effects among 25 genes and predict RNA trajectories under 24 new, two-stage intervention protocols. Separate direct regulator-to-target effects from indirect responses caused by shared translation and delayed protein turnover. Use all supplied calibration data, including the intervention mapping, to assess the fitted model and explain uncertainty.
|
| 6 |
+
|
| 7 |
+
## Mechanistic model (B1)
|
| 8 |
+
|
| 9 |
+
For each gene `i`, let `x_i` be RNA abundance and `p_i` protein abundance. The signed direct-edge matrix `W` uses `[target, source]` indexing: `W[i,j] > 0` means protein `j` activates transcription of `i`; `W[i,j] < 0` means it represses transcription; `W[i,j] = 0` means no direct edge. The diagonal is zero. The reference network is sparse, with at most three direct regulators per target. Infer `W`; do not treat global translation competition as a direct edge.
|
| 10 |
+
|
| 11 |
+
Total RNA load limits translation. Define `q_i(x)=12*x_i/(5+sum_j x_j)` and `H(p_i)=p_i^3/(0.65^3+p_i^3)`. For direct parent `j` of target `i`, set `f_ij=H(p_j)` for activation and `f_ij=1-H(p_j)` for repression; set it to zero for an absent or inactivated parent. With gene-specific positive basal transcription `b_i`, nonnegative target-specific cooperative strength `kappa_i`, RNA decay `gamma_m_i`, and protein decay `gamma_p_i`, the equations are:
|
| 12 |
+
|
| 13 |
+
```text
|
| 14 |
+
dp_i/dt = gamma_p_i * (q_i(x) - p_i)
|
| 15 |
+
dx_i/dt = b_i
|
| 16 |
+
+ sum_{j: W[i,j] != 0} abs(W[i,j]) * f_ij
|
| 17 |
+
+ kappa_i * sum_{j<k} f_ij * f_ik
|
| 18 |
+
- gamma_m_i * x_i
|
| 19 |
+
```
|
| 20 |
+
|
| 21 |
+
The interaction sum is over unordered pairs of distinct direct regulators. An inactivated regulator has zero activity in both terms, removing its main contribution and every interaction involving it; its own RNA and protein still evolve. The signed direct effects remain monotone in their activating or repressing direction. An RNA clamp sets that `x_i` immediately at the segment boundary and holds it there; it does not reset `p_i`. Every segment's intervention lists replace the previous lists. Releasing an intervention leaves all internal states continuous.
|
| 22 |
+
|
| 23 |
+
## Suggested reference inference (B1)
|
| 24 |
+
|
| 25 |
+
The following is one fast, reproducible route. It is guidance rather than an additional scoring rule.
|
| 26 |
+
|
| 27 |
+
1. Estimate each `gamma_m_i` and `gamma_p_i` from its noise-free retention series `R_i(t)`. Fit `log(R_i(t))=-gamma_i*t` through the origin by least squares: `gamma_i = -sum(t*log(R_i(t)))/sum(t*t)`.
|
| 28 |
+
2. Smooth each of the eight RNA time courses with a cubic Savitzky–Golay filter using a 9-sample window. Starting from its measured `initial_protein`, reconstruct protein using the first-order protein equation. On interval `k`, let `u_k=q(x_k)`, `s=(u_{k+1}-u_k)/dt`, and `e=exp(-gamma_p*dt)`; use `p_{k+1}=p_k*e+u_k*(1-e)+s*(dt-(1-e)/gamma_p)`.
|
| 29 |
+
3. Evaluate `H(p)` and its pairwise products, and form trapezoidal cumulative integrals of `H`, `H_j*H_k`, and `x`. For each trajectory, use windows of eight sample intervals, starting every four intervals. For target `i`, a window supplies `A_j=mean_window(H(p_j))`, `B_jk=mean_window(H(p_j)*H(p_k))`, and `Y_i=(delta x_i + gamma_m_i*integral_window(x_i))/window_duration`.
|
| 30 |
+
4. Retain all 24 non-self candidate regulators for each target. Enumerate every two- and three-parent subset and every activating/repressing sign assignment; do not fix signs from an additive screen. For each model, build columns for the intercept, each mean `f_j`, and the mean of `sum_{j<k} f_j*f_k`. The pairwise means follow from `A` and `B` (for example, an activating/repressing pair has mean `A_j-B_jk`). Fit nonnegative intercept, parent magnitudes, and cooperative strength with NNLS. Rank fits with `BIC=n*log(RSS/n+1e-20)+2*(m+2)*log(n)`, where `m` is the number of parents. Batched unconstrained least squares can supply RSS lower bounds to skip models that cannot improve the current constrained fit; this does not discard any potentially better support or sign assignment. Set absent-edge weights exactly to zero.
|
| 31 |
+
5. Refine the continuous parameters after structure selection. Interpolate the smoothed RNA with a cubic spline, reconstruct protein with the known ODE on a grid eight times finer than the observation grid, and compute each signed promoter activity and cooperative pair sum. Fit the original noisy RNA levels to the exponential response of the RNA equation using nonnegative least squares, with each replicate's initial RNA as an additional nuisance parameter. For production feature `u(t)`, its convolution obeys `c'=u-gamma_m*c`, `c(0)=0`; propagate it exactly for piecewise-linear `u`. Weight RNA residuals by `1/(0.2+smoothed_RNA)`. This avoids coarse-window integration bias and differencing noise in the final parameter estimates.
|
| 32 |
+
6. Integrate the coupled 50-state RNA/protein system under each listed calibration intervention and each two-segment query using the fitted `W`, `b`, and `kappa`. Carry RNA and protein across segment boundaries, except that any RNA named in the new segment's clamps is reset to its listed value; proteins are never reset. Initialize each query with `late_rna[0]` and `wildtype_protein`; sample at the 12 specified times. Use the 15 calibration outcomes to check transfer behavior and discuss any mismatch.
|
| 33 |
+
|
| 34 |
+
## Inputs
|
| 35 |
+
|
| 36 |
+
All files are in `data/`; every listed key is required. No input field is optional or nullable, and there are no YAML inputs.
|
| 37 |
+
|
| 38 |
+
- `study.json`: integer `schema_version`, integer `n_genes` (25), strings `time_unit`, `rna_units`, `protein_units`, `measurements`, `decay_assays`, `interventions`, `prediction`, `direct_edge`, `biology`, and numeric `observation_noise_relative`. Use its definitions for units, measurements, interventions, the meaning of a direct edge, and the exact kinetic laws in `biology`. These laws and their constants are public at all prompt levels.
|
| 39 |
+
- `observations.npz`: NumPy NPZ archive containing:
|
| 40 |
+
- `times`: shape `(81,)`, dtype `float64`.
|
| 41 |
+
- `rna`: shape `(8,81,25)`, dtype `float64`.
|
| 42 |
+
- `initial_protein`: shape `(8,25)`, dtype `float64`.
|
| 43 |
+
- `decay_times`: shape `(6,)`, dtype `float64`.
|
| 44 |
+
- `rna_retention`: shape `(25,6)`, dtype `float64`.
|
| 45 |
+
- `protein_retention`: shape `(25,6)`, dtype `float64`.
|
| 46 |
+
- `late_rna`: shape `(16,25)`, dtype `float64`.
|
| 47 |
+
- `wildtype_protein`: shape `(25,)`, dtype `float64`.
|
| 48 |
+
|
| 49 |
+
Eight no-intervention RNA trajectories cover 0–4 minutes. RNA noise is independent normal noise with standard deviation `0.003*(0.2+true_RNA)`, clipped below at zero. Initial proteins, the untreated reference, and retention measurements are noise free. The retention arrays come from independent synthesis-blocking decay assays. `late_rna[0]` is the untreated reference; rows 1–15 are the measured RNA values 45 minutes after their listed calibration intervention and are not guaranteed to be steady states.
|
| 50 |
+
- `experiments.json`: `calibration_interventions` is a 15-object list, in row order for `late_rna[1:16]`; `query_protocols` is a 24-object list. Every calibration object has a string `id`, an `rna_clamps` list, and an `inactive_regulators` list. Every query object has a string `id` and a two-element `segments` list. Each segment has a numeric `duration` of 1.5 minutes, an `rna_clamps` list, and an `inactive_regulators` list. A clamp is an object with integer `gene` 0–24 and nonnegative numeric `value`; an inactive-regulator list contains integer gene IDs 0–24. All intervention targets and operation types are supplied at every prompt level.
|
| 51 |
+
|
| 52 |
+
Queries start independently from `late_rna[0]` and `wildtype_protein`. Within each 1.5-minute segment, the listed controls apply for the whole segment. When a control is absent from the next segment, it is released without resetting RNA or protein; newly clamped RNA is set to its listed value at the boundary, while protein is carried forward. Record six evenly spaced RNA samples in each segment, at `duration/6, 2*duration/6, ..., duration` relative to that segment.
|
| 53 |
+
|
| 54 |
+
## Outputs and evaluation
|
| 55 |
+
|
| 56 |
+
Write `analysis.py`, `report.md`, and `predictions.npz` at the workspace root. The NPZ must contain finite arrays `weights` `(25,25)` `float64`, `future` `(24,12,25)` `float64`, and `hubs` `(2,)` `int64`. `weights[i,j]` is the signed direct effect of source `j` on target `i`; keep its diagonal zero and set absent edges to exactly zero. The scorer treats `abs(weight)<=1e-6` as zero. `future[q,t,i]` is the RNA prediction at the ordered sample time for query `q`, gene `i`. `hubs` contains two distinct gene IDs in 0–24, ranked by direct out-degree (ties go to lower ID).
|
| 57 |
+
|
| 58 |
+
The raw 100-point score assigns 60 to signed direct-edge F1, 20 to 24 prospective forecasts, 10 to hub overlap, and 10 to chance-adjusted average precision of edge ranking by `abs(weights)`. A forecast succeeds when its RMSE, normalized per gene by `0.2 + mean untreated-reference RNA`, is at most 0.10; forecast credit is proportional to the fraction of successful queries. The scorer passes at 90 points. Explain how the identified direct network accounts for the observations and predictions, and state a limitation of this synthetic model.
|
| 59 |
+
|
| 60 |
+
### Reliability requirement
|
| 61 |
+
|
| 62 |
+
This study claims both reliable direct-network recovery and reliable intervention transfer. The raw 60/20/10/10 score above remains visible. Two additional, predeclared reliability checks determine the final score:
|
| 63 |
+
|
| 64 |
+
- Network reliability: signed direct-edge F1 is at least 0.90.
|
| 65 |
+
- Transfer reliability: at least 90% of the 24 query protocols have **every** gene and sampled time within normalized absolute error 0.03. For protocol q this error is `max_{t,i} abs(future[q,t,i]-truth[q,t,i])/(0.2+reference_RNA[i])`; `reference_RNA` is `late_rna[0]`. Thus at least 22 protocols must meet this componentwise tolerance. This is stricter than the raw forecast component's mean-square criterion.
|
| 66 |
+
|
| 67 |
+
If both checks pass, the final score equals the raw score. If exactly one passes, the final score is `min(raw_score,35)`. If neither passes, it is `min(raw_score,30)`. All levels use the same thresholds; the scorer never reads the prompt level. The score report retains raw points, both reliability measures, the applied cap, and the final score. A fully correct solution earns 100 at any prompt level.
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/prompt_b2.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Recovering a Gene Network from Mixed Perturbations
|
| 2 |
+
|
| 3 |
+
## Scientific goal
|
| 4 |
+
|
| 5 |
+
Infer the signed **direct** transcriptional effects among 25 genes and predict RNA trajectories under 24 new, two-stage intervention protocols. Separate direct regulator-to-target effects from indirect responses caused by shared translation and delayed protein turnover. Use all supplied calibration data, including the intervention mapping, to assess the fitted model and explain uncertainty.
|
| 6 |
+
|
| 7 |
+
## Mechanistic model (B2)
|
| 8 |
+
|
| 9 |
+
For each gene `i`, let `x_i` be RNA abundance and `p_i` protein abundance. The signed direct-edge matrix `W` uses `[target, source]` indexing: `W[i,j] > 0` means protein `j` activates transcription of `i`; `W[i,j] < 0` means it represses transcription of `i`; `W[i,j] = 0` means no direct edge. The diagonal is zero. The reference network is sparse, with at most three direct regulators per target. Infer `W`; do not treat global translation competition as a direct edge.
|
| 10 |
+
|
| 11 |
+
The coupled mechanism has a shared, saturating translation input, delayed protein-mediated transcription, and target-specific cooperative promoter occupancy. Define `q_i(x)=12*x_i/(5+sum_j x_j)` and `H(p_i)=p_i^3/(0.65^3+p_i^3)`. For a direct parent `j` of target `i`, set `f_ij=H(p_j)` for activation and `f_ij=1-H(p_j)` for repression; set it to zero for an absent or inactivated parent. With gene-specific positive basal transcription `b_i`, nonnegative target-specific cooperative strength `kappa_i`, RNA decay `gamma_m_i`, and protein decay `gamma_p_i`, the equations are:
|
| 12 |
+
|
| 13 |
+
```text
|
| 14 |
+
dp_i/dt = gamma_p_i * (q_i(x) - p_i)
|
| 15 |
+
dx_i/dt = b_i
|
| 16 |
+
+ sum_{j: W[i,j] != 0} abs(W[i,j]) * f_ij
|
| 17 |
+
+ kappa_i * sum_{j<k} f_ij * f_ik
|
| 18 |
+
- gamma_m_i * x_i
|
| 19 |
+
```
|
| 20 |
+
|
| 21 |
+
The interaction sum is over unordered pairs of distinct direct regulators. An inactivated regulator has zero activity in both terms, removing its main contribution and every interaction involving it; its own RNA and protein still evolve. The signed direct effects remain monotone in their activating or repressing direction. An RNA clamp sets that `x_i` immediately at the segment boundary and holds it there, without resetting its protein. Each segment's intervention lists replace the prior segment's lists. Releasing an intervention leaves all internal states continuous.
|
| 22 |
+
|
| 23 |
+
The model family and intervention mechanisms are provided, but no numeric estimation recipe is prescribed. Choose an inference approach, state its assumptions, and use the RNA time courses, independent turnover assays, reference condition, and fully mapped calibration interventions to infer the network and check its prospective predictions. Account for the fact that RNA and protein need not equilibrate together and that shared translation can create indirect responses even when no direct edge exists.
|
| 24 |
+
|
| 25 |
+
## Inputs
|
| 26 |
+
|
| 27 |
+
All files are in `data/`; every listed key is required. No input field is optional or nullable, and there are no YAML inputs.
|
| 28 |
+
|
| 29 |
+
- `study.json`: integer `schema_version`, integer `n_genes` (25), strings `time_unit`, `rna_units`, `protein_units`, `measurements`, `decay_assays`, `interventions`, `prediction`, `direct_edge`, `biology`, and numeric `observation_noise_relative`. Use its definitions for units, measurements, interventions, the meaning of a direct edge, and the exact kinetic laws in `biology`. These laws and their constants are public at all prompt levels.
|
| 30 |
+
- `observations.npz`: NumPy NPZ archive containing:
|
| 31 |
+
- `times`: shape `(81,)`, dtype `float64`.
|
| 32 |
+
- `rna`: shape `(8,81,25)`, dtype `float64`.
|
| 33 |
+
- `initial_protein`: shape `(8,25)`, dtype `float64`.
|
| 34 |
+
- `decay_times`: shape `(6,)`, dtype `float64`.
|
| 35 |
+
- `rna_retention`: shape `(25,6)`, dtype `float64`.
|
| 36 |
+
- `protein_retention`: shape `(25,6)`, dtype `float64`.
|
| 37 |
+
- `late_rna`: shape `(16,25)`, dtype `float64`.
|
| 38 |
+
- `wildtype_protein`: shape `(25,)`, dtype `float64`.
|
| 39 |
+
|
| 40 |
+
Eight no-intervention RNA trajectories cover 0–4 minutes. RNA noise is independent normal noise with standard deviation `0.003*(0.2+true_RNA)`, clipped below at zero. Initial proteins, the untreated reference, and retention measurements are noise free. The retention arrays come from independent synthesis-blocking decay assays. `late_rna[0]` is the untreated reference; rows 1–15 are the measured RNA values 45 minutes after their listed calibration intervention and are not guaranteed to be steady states.
|
| 41 |
+
- `experiments.json`: `calibration_interventions` is a 15-object list, in row order for `late_rna[1:16]`; `query_protocols` is a 24-object list. Every calibration object has a string `id`, an `rna_clamps` list, and an `inactive_regulators` list. Every query object has a string `id` and a two-element `segments` list. Each segment has a numeric `duration` of 1.5 minutes, an `rna_clamps` list, and an `inactive_regulators` list. A clamp is an object with integer `gene` 0–24 and nonnegative numeric `value`; an inactive-regulator list contains integer gene IDs 0–24. All intervention targets and operation types are supplied at every prompt level.
|
| 42 |
+
|
| 43 |
+
Queries start independently from `late_rna[0]` and `wildtype_protein`. Within each 1.5-minute segment, the listed controls apply for the whole segment. When a control is absent from the next segment, it is released without resetting RNA or protein; newly clamped RNA is set to its listed value at the boundary, while protein is carried forward. Record six evenly spaced RNA samples in each segment, at `duration/6, 2*duration/6, ..., duration` relative to that segment.
|
| 44 |
+
|
| 45 |
+
## Outputs and evaluation
|
| 46 |
+
|
| 47 |
+
Write `analysis.py`, `report.md`, and `predictions.npz` at the workspace root. The NPZ must contain finite arrays `weights` `(25,25)` `float64`, `future` `(24,12,25)` `float64`, and `hubs` `(2,)` `int64`. `weights[i,j]` is the signed direct effect of source `j` on target `i`; keep its diagonal zero and set absent edges to exactly zero. The scorer treats `abs(weight)<=1e-6` as zero. `future[q,t,i]` is the RNA prediction at the ordered sample time for query `q`, gene `i`. `hubs` contains two distinct gene IDs in 0–24, ranked by direct out-degree (ties go to lower ID).
|
| 48 |
+
|
| 49 |
+
The raw 100-point score assigns 60 to signed direct-edge F1, 20 to 24 prospective forecasts, 10 to hub overlap, and 10 to chance-adjusted average precision of edge ranking by `abs(weights)`. A forecast succeeds when its RMSE, normalized per gene by `0.2 + mean untreated-reference RNA`, is at most 0.10; forecast credit is proportional to the fraction of successful queries. The scorer passes at 90 points. Explain how the identified direct network accounts for the observations and predictions, and state a limitation of this synthetic model.
|
| 50 |
+
|
| 51 |
+
### Reliability requirement
|
| 52 |
+
|
| 53 |
+
This study claims both reliable direct-network recovery and reliable intervention transfer. The raw 60/20/10/10 score above remains visible. Two additional, predeclared reliability checks determine the final score:
|
| 54 |
+
|
| 55 |
+
- Network reliability: signed direct-edge F1 is at least 0.90.
|
| 56 |
+
- Transfer reliability: at least 90% of the 24 query protocols have **every** gene and sampled time within normalized absolute error 0.03. For protocol q this error is `max_{t,i} abs(future[q,t,i]-truth[q,t,i])/(0.2+reference_RNA[i])`; `reference_RNA` is `late_rna[0]`. Thus at least 22 protocols must meet this componentwise tolerance. This is stricter than the raw forecast component's mean-square criterion.
|
| 57 |
+
|
| 58 |
+
If both checks pass, the final score equals the raw score. If exactly one passes, the final score is `min(raw_score,35)`. If neither passes, it is `min(raw_score,30)`. All levels use the same thresholds; the scorer never reads the prompt level. The score report retains raw points, both reliability measures, the applied cap, and the final score. A fully correct solution earns 100 at any prompt level.
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/prompt_b3.md
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Recovering a Gene Network from Mixed Perturbations
|
| 2 |
+
|
| 3 |
+
## Scientific goal
|
| 4 |
+
|
| 5 |
+
Infer signed **direct** transcriptional effects among 25 genes and predict RNA trajectories under 24 new, two-stage intervention protocols. Separate direct regulator-to-target effects from indirect responses that can arise through shared translation and delayed protein turnover. Use all supplied calibration data, including the intervention mapping, to assess the fitted model and explain uncertainty.
|
| 6 |
+
|
| 7 |
+
The circuit is a controlled synthetic kinetic system. RNA and functional protein are distinct measured quantities; transcription has gene-specific basal production and turnover, and protein turnover is measured independently. The exact functional laws of the regulatory and shared-resource dynamics, including the Hill order and threshold, are specified in `study.json` and are shared across experiments. The direct wiring and gene-specific numerical parameters remain unknown and must be inferred from the observations. Direct transcriptional effects are signed. Infer the direct network from the observations rather than treating every correlated response as a direct edge. The study definitions and intervention records are part of the input.
|
| 8 |
+
|
| 9 |
+
## Inputs
|
| 10 |
+
|
| 11 |
+
All files are in `data/`; every listed key is required. No input field is optional or nullable, and there are no YAML inputs.
|
| 12 |
+
|
| 13 |
+
- `study.json`: integer `schema_version`, integer `n_genes` (25), strings `time_unit`, `rna_units`, `protein_units`, `measurements`, `decay_assays`, `interventions`, `prediction`, `direct_edge`, `biology`, and numeric `observation_noise_relative`. Use its definitions for units, measurements, interventions, the meaning of a direct edge, and the exact kinetic laws in `biology`. These laws and their constants are public at all prompt levels.
|
| 14 |
+
- `observations.npz`: NumPy NPZ archive containing:
|
| 15 |
+
- `times`: shape `(81,)`, dtype `float64`.
|
| 16 |
+
- `rna`: shape `(8,81,25)`, dtype `float64`.
|
| 17 |
+
- `initial_protein`: shape `(8,25)`, dtype `float64`.
|
| 18 |
+
- `decay_times`: shape `(6,)`, dtype `float64`.
|
| 19 |
+
- `rna_retention`: shape `(25,6)`, dtype `float64`.
|
| 20 |
+
- `protein_retention`: shape `(25,6)`, dtype `float64`.
|
| 21 |
+
- `late_rna`: shape `(16,25)`, dtype `float64`.
|
| 22 |
+
- `wildtype_protein`: shape `(25,)`, dtype `float64`.
|
| 23 |
+
|
| 24 |
+
Eight no-intervention RNA trajectories cover 0–4 minutes. RNA noise is independent normal noise with standard deviation `0.003*(0.2+true_RNA)`, clipped below at zero. Initial proteins, the untreated reference, and retention measurements are noise free. The retention arrays come from independent synthesis-blocking decay assays. `late_rna[0]` is the untreated reference; rows 1–15 are the measured RNA values 45 minutes after their listed calibration intervention and are not guaranteed to be steady states.
|
| 25 |
+
- `experiments.json`: `calibration_interventions` is a 15-object list, in row order for `late_rna[1:16]`; `query_protocols` is a 24-object list. Every calibration object has a string `id`, an `rna_clamps` list, and an `inactive_regulators` list. Every query object has a string `id` and a two-element `segments` list. Each segment has a numeric `duration` of 1.5 minutes, an `rna_clamps` list, and an `inactive_regulators` list. A clamp is an object with integer `gene` 0–24 and nonnegative numeric `value`; an inactive-regulator list contains integer gene IDs 0–24. All intervention targets and operation types are supplied at every prompt level.
|
| 26 |
+
|
| 27 |
+
Queries start independently from `late_rna[0]` and `wildtype_protein`. Within each 1.5-minute segment, the listed controls apply for the whole segment. An RNA clamp instantaneously sets the specified RNA at the segment boundary and holds it there; it does not reset protein. For each listed inactive regulator, remove all of its promoter action during that segment while its own RNA and protein continue to evolve. Each segment lists the complete current controls; controls absent from the next segment are released without resetting internal states, and newly clamped RNA is set to its listed value at the boundary. Record six evenly spaced RNA samples in each segment, at `duration/6, 2*duration/6, ..., duration` relative to that segment.
|
| 28 |
+
|
| 29 |
+
## Outputs
|
| 30 |
+
|
| 31 |
+
Write `analysis.py`, `report.md`, and `predictions.npz` at the workspace root. The NPZ must contain finite arrays `weights` `(25,25)` `float64`, `future` `(24,12,25)` `float64`, and `hubs` `(2,)` `int64`. `weights[i,j]` is the signed direct effect of source `j` on target `i`; keep its diagonal zero and set absent edges to exactly zero. The scorer treats `abs(weight)<=1e-6` as zero. `future[q,t,i]` is the RNA prediction at the ordered sample time for query `q`, gene `i`. `hubs` contains two distinct gene IDs in 0–24, ranked by direct out-degree (ties go to lower ID).
|
| 32 |
+
|
| 33 |
+
The raw 100-point score assigns 60 to signed direct-edge F1, 20 to 24 prospective forecasts, 10 to hub overlap, and 10 to chance-adjusted average precision of edge ranking by `abs(weights)`. A forecast succeeds when its RMSE, normalized per gene by `0.2 + mean untreated-reference RNA`, is at most 0.10; forecast credit is proportional to the fraction of successful queries. The scorer passes at 90 points. Explain how the identified direct network accounts for the observations and predictions, describe your inference and validation approach, and state a limitation of this synthetic model.
|
| 34 |
+
|
| 35 |
+
### Reliability requirement
|
| 36 |
+
|
| 37 |
+
This study claims both reliable direct-network recovery and reliable intervention transfer. The raw 60/20/10/10 score above remains visible. Two additional, predeclared reliability checks determine the final score:
|
| 38 |
+
|
| 39 |
+
- Network reliability: signed direct-edge F1 is at least 0.90.
|
| 40 |
+
- Transfer reliability: at least 90% of the 24 query protocols have **every** gene and sampled time within normalized absolute error 0.03. For protocol q this error is `max_{t,i} abs(future[q,t,i]-truth[q,t,i])/(0.2+reference_RNA[i])`; `reference_RNA` is `late_rna[0]`. Thus at least 22 protocols must meet this componentwise tolerance. This is stricter than the raw forecast component's mean-square criterion.
|
| 41 |
+
|
| 42 |
+
If both checks pass, the final score equals the raw score. If exactly one passes, the final score is `min(raw_score,35)`. If neither passes, it is `min(raw_score,30)`. All levels use the same thresholds; the scorer never reads the prompt level. The score report retains raw points, both reliability measures, the applied cap, and the final score. A fully correct solution earns 100 at any prompt level.
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/prompt_b4.md
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Recovering a Gene Network from Mixed Perturbations
|
| 2 |
+
|
| 3 |
+
## Scientific goal
|
| 4 |
+
|
| 5 |
+
Infer signed **direct** transcriptional effects among 25 genes and predict RNA trajectories under 24 new, two-stage intervention protocols. Separate direct regulator-to-target effects from indirect responses that can arise through shared translation and delayed protein turnover. Use all supplied calibration data, including the intervention mapping, to assess the fitted model and explain uncertainty.
|
| 6 |
+
|
| 7 |
+
The circuit is a controlled synthetic kinetic system. RNA and functional protein are distinct measured quantities; transcription has gene-specific basal production and turnover, and protein turnover is measured independently. The exact functional laws of the regulatory and shared-resource dynamics, including the Hill order and threshold, are specified in `study.json` and are shared across experiments. The direct wiring and gene-specific numerical parameters remain unknown and must be inferred from the observations. Direct transcriptional effects are signed. Infer the direct network from the observations rather than treating every correlated response as a direct edge. The study definitions and intervention records are part of the input.
|
| 8 |
+
|
| 9 |
+
## Inputs
|
| 10 |
+
|
| 11 |
+
All files are in `data/`; every listed key is required. No input field is optional or nullable, and there are no YAML inputs.
|
| 12 |
+
|
| 13 |
+
- `study.json`: integer `schema_version`, integer `n_genes` (25), strings `time_unit`, `rna_units`, `protein_units`, `measurements`, `decay_assays`, `interventions`, `prediction`, `direct_edge`, `biology`, and numeric `observation_noise_relative`. Use its definitions for units, measurements, interventions, the meaning of a direct edge, and the exact kinetic laws in `biology`. These laws and their constants are public at all prompt levels.
|
| 14 |
+
- `observations.npz`: NumPy NPZ archive containing:
|
| 15 |
+
- `times`: shape `(81,)`, dtype `float64`.
|
| 16 |
+
- `rna`: shape `(8,81,25)`, dtype `float64`.
|
| 17 |
+
- `initial_protein`: shape `(8,25)`, dtype `float64`.
|
| 18 |
+
- `decay_times`: shape `(6,)`, dtype `float64`.
|
| 19 |
+
- `rna_retention`: shape `(25,6)`, dtype `float64`.
|
| 20 |
+
- `protein_retention`: shape `(25,6)`, dtype `float64`.
|
| 21 |
+
- `late_rna`: shape `(16,25)`, dtype `float64`.
|
| 22 |
+
- `wildtype_protein`: shape `(25,)`, dtype `float64`.
|
| 23 |
+
|
| 24 |
+
Eight no-intervention RNA trajectories cover 0–4 minutes. RNA noise is independent normal noise with standard deviation `0.003*(0.2+true_RNA)`, clipped below at zero. Initial proteins, the untreated reference, and retention measurements are noise free. The retention arrays come from independent synthesis-blocking decay assays. `late_rna[0]` is the untreated reference; rows 1–15 are the measured RNA values 45 minutes after their listed calibration intervention and are not guaranteed to be steady states.
|
| 25 |
+
- `experiments.json`: `calibration_interventions` is a 15-object list, in row order for `late_rna[1:16]`; `query_protocols` is a 24-object list. Every calibration object has a string `id`, an `rna_clamps` list, and an `inactive_regulators` list. Every query object has a string `id` and a two-element `segments` list. Each segment has a numeric `duration` of 1.5 minutes, an `rna_clamps` list, and an `inactive_regulators` list. A clamp is an object with integer `gene` 0–24 and nonnegative numeric `value`; an inactive-regulator list contains integer gene IDs 0–24. All intervention targets and operation types are supplied at every prompt level.
|
| 26 |
+
|
| 27 |
+
Queries start independently from `late_rna[0]` and `wildtype_protein`. Within each 1.5-minute segment, the listed controls apply for the whole segment. An RNA clamp instantaneously sets the specified RNA at the segment boundary and holds it there; it does not reset protein. For each listed inactive regulator, remove all of its promoter action during that segment while its own RNA and protein continue to evolve. Each segment lists the complete current controls; controls absent from the next segment are released without resetting internal states, and newly clamped RNA is set to its listed value at the boundary. Record six evenly spaced RNA samples in each segment, at `duration/6, 2*duration/6, ..., duration` relative to that segment.
|
| 28 |
+
|
| 29 |
+
## Outputs
|
| 30 |
+
|
| 31 |
+
Write `analysis.py`, `report.md`, and `predictions.npz` at the workspace root. The NPZ must contain finite arrays `weights` `(25,25)` `float64`, `future` `(24,12,25)` `float64`, and `hubs` `(2,)` `int64`. `weights[i,j]` is the signed direct effect of source `j` on target `i`; keep its diagonal zero and set absent edges to exactly zero. The scorer treats `abs(weight)<=1e-6` as zero. `future[q,t,i]` is the RNA prediction at the ordered sample time for query `q`, gene `i`. `hubs` contains two distinct gene IDs in 0–24, ranked by direct out-degree (ties go to lower ID).
|
| 32 |
+
|
| 33 |
+
The raw 100-point score assigns 60 to signed direct-edge F1, 20 to 24 prospective forecasts, 10 to hub overlap, and 10 to chance-adjusted average precision of edge ranking by `abs(weights)`. A forecast succeeds when its RMSE, normalized per gene by `0.2 + mean untreated-reference RNA`, is at most 0.10; forecast credit is proportional to the fraction of successful queries. The scorer passes at 90 points. Explain how the identified direct network accounts for the observations and predictions, describe your inference and validation approach, and state a limitation of this synthetic model.
|
| 34 |
+
|
| 35 |
+
### Reliability requirement
|
| 36 |
+
|
| 37 |
+
This study claims both reliable direct-network recovery and reliable intervention transfer. The raw 60/20/10/10 score above remains visible. Two additional, predeclared reliability checks determine the final score:
|
| 38 |
+
|
| 39 |
+
- Network reliability: signed direct-edge F1 is at least 0.90.
|
| 40 |
+
- Transfer reliability: at least 90% of the 24 query protocols have **every** gene and sampled time within normalized absolute error 0.03. For protocol q this error is `max_{t,i} abs(future[q,t,i]-truth[q,t,i])/(0.2+reference_RNA[i])`; `reference_RNA` is `late_rna[0]`. Thus at least 22 protocols must meet this componentwise tolerance. This is stricter than the raw forecast component's mean-square criterion.
|
| 41 |
+
|
| 42 |
+
If both checks pass, the final score equals the raw score. If exactly one passes, the final score is `min(raw_score,35)`. If neither passes, it is `min(raw_score,30)`. All levels use the same thresholds; the scorer never reads the prompt level. The score report retains raw points, both reliability measures, the applied cap, and the final score. A fully correct solution earns 100 at any prompt level.
|
| 43 |
+
|
| 44 |
+
## Additional laboratory notes
|
| 45 |
+
|
| 46 |
+
The reported time coordinate is in minutes; if you use a solver that expects seconds, convert all times and rate constants consistently. The order of replicate trajectories in `rna` is arbitrary and carries no batch or treatment meaning. Reordering query records changes only the corresponding output ordering. A consistent permutation of gene identifiers permutes the rows and columns of a direct-effect matrix and preserves its number of nonzero entries.
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/reference/ground_truth.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:36913f2fcae4929be374aabf21bff6c07431c45819d9148d53d9e5fb6e5f87c9
|
| 3 |
+
size 65624
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/reference/predictions.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a45ab4da175adf7a9caade0112e7d6bc70ee4d3176e738adac8a290c3e2fad69
|
| 3 |
+
size 63368
|
tasks/biology.mixed_perturbation_gene_transfer__seed31415/reference/report.md
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
Original forward-model ground truth. The separate reference solution estimates the parameters from public measurements.
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/data/promoters.fa
ADDED
|
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
>seq_00
|
| 2 |
+
GTGTGACTGCTTACAATGGCGCGGAGTAAATGGGAAGGCCAGCGGGCTGGGGCACAAGATTGCCGTGTAACTGGATCACCGTCCCCTTATCGTAGGACCGCCGCTCCTATTTCCACGTCACAGAGCTTAGGTCTCGAAGTATGGATCGCACAAAGTTTTAGTCACTTCCCCTTGCAAAAGGATTGTGCGAGTTTGGGACTTTTCGCGGTCAGTCCATAGATACTAGTCTTTGTATAGAATCAAGCAAAGATCACATATGTAGACGACTCCTTAATCTCCGCCTCGTAGAACCAGACCAGGGAGAGCCGTCTCTATTAGGGTTACATATTTAATAAATGCAGTTCAAAGCAAGTGGGCTTTTTCGGGAGACATCTAGATATAAAAACGTACAGAGTTCAGTGCGCCAAAGCACGAATCCTTGGGGGCTATGTATTCCGCCTATGCAACGTGTTGCCCATCATAGGGGTATAACACTCCACGTTGATAGGGGACGTACATTACACGTTAACAGAAAATCACGACAGGGATGATATCTCAAGCTCGCTGGGATCAAGCGATCATCTCCACCTTGTAAGTCTTCTTTGCTACAAGCTGTATGGGGAACCTTAGCACGTTCTAAGCCTCGTATGCTTCACCGCGACTATCGGCACTCTTAGTACATAAAACTTGTGACCAAACTATCGTATACCTGCGTACAGCCCTACCCGTTCCTCGCCCATCCGGGCAAATTGGGAATTTCGCCAACACAATGGCTCCCCATACCAATGTACGATCGCAACGGTCGGCTTAGATTCTAATTTGGTAATAACGTCCAGTGCAGTCCTCGTATTATCCTGCTCACGCTGCAGGCCGAGTCCTACACCGTCCCTGCCATCGTACATAATTGTTCTGTCGAGCTTGCGGAACGTGTGCTTTCGACTACGAGACAAGCTTATGCTATCGTCATATATGGTATGGGGAAGAAAGGCGGTCTTCAGACGTTGGCCAAGGCTCAGATGCC
|
| 3 |
+
>seq_01
|
| 4 |
+
TGAAGATGTCTACCGGGGACCTCGGCTTCCCCTGTTTGGCTCGAGGTTAGGTCCAAAGAAATTCTGACAAGATTAGCCATACATTGTTCTGGAGGTAAAAAATAATCAGCCCAAAATGGTGCATCACGACCCCCCGCTTGTCGCTATCCCTTTGCGGAGACTATGGCAAGGGCAAATCATGACCTCGTCCAGAAATCGAAGAGCTACGAGACTTATCACTCCGAGGATTCGCCGGAAGGCTCTCCAAGTACAAGGAAGTTCTCATCTGTTGAGGGAGGCCGAGTCTTCGTTACCTACATAGCAGACGGGAGCACTCTATGGTGAATGATGAAGTACGTGTAGGATCACTGTCTGCATATCGCATGAGTAAGCCATACAGACTATTGACTACTGCGGCATGGAGTGTGTTGGTGTGGTAAACAGTGAGGCTCCGATGGAAGTTGGGGGTGCGGGTTGGAGATAGGGAATGACCCGTTGACAGGTAGATTCGCGATCCTGAGGAGGCGGTCCCGGTCGGAAATGCGTGCCAGAACGCGTCTCGGTCTAAAACCGTGTGCCGAGAAAGGATTGACACGAGGACTCGCCAACGAGACGGTGTTCTGACGGTCGCCGATGTGTTAGAATATAGCTCCCGTGGCTAGCGTACGAATCATACGGACGTGCCCCTGACGGCTGTTCCATGGGGTGGACGAGTAGTGTACATCGTAATCCCTATAATTAGAATAGAATCCTGAGGTGTGCCACAGTCATGGGTCGGGCACACGGTTTGCATAGGTAATTTATAGGCTGGATCGGAGGCATTGGCAGGTCCGAGATATTGCCTATCTTCGCAACAGGCACGGTGAGTCCTTCGCCCGAAGTATCTAGTAGCGCATGTGACGCGTGGAATCGCCCCCAAGTAACTACGATAGTTTGGTTCCAGGAGGGTACCTAAAATGACGCAGCTAATAACTAAATCGGAAGCCTATGGGTTGACACTCACAACGCCTGGCGTGGCCTC
|
| 5 |
+
>seq_02
|
| 6 |
+
ATGCTAGGGTTCCCACGTGGCACCCCAGGCACTTAAGGCCACTTGCCATCGGGAGGCACCATATGGGTCTTTCGACAGCGTCCCTTAGAACCCAATTACTAGGGGGGACGATAAATTATCGCGATGGTTATAAGATCTCACCGCATGGCAGGATAACTACCCGGAAAGACACCGCGCCAATGCCATCCGGCCATATTATGCAACGGTCCCGTCTATGGGGGTCATCTCGGCCGGTATGTTTGCAGTTTCTTGTTAAACCCTAGTGTACGAGTTGCCGGTAGACTCACGTGCCCACCCACTCATGGCGTCTTGGCGGATTGAGTAACTGTTGAGGAGATTGTGCGCGTTCAAGGTGTAGATATAACGGGCGCCCGCTTCCTTTGCGTGAGAAAAGCTTACGTCGGAACGGGTTAAACTGGGACCATGGACGAACAGATCGAGGGCACGTAAGAGATTTCCTGGCATGTAGACATGGCCGCCAATTCCGCGACGCTGTTGACCCTCTCAGGGTCCAGTCAGTCAAGTCATACGATTGGTCGCGACCTGATAGCAATGACACGATCGCGTAGTCAGTGAAGGGCCGGGTCGAATATGCTGCCTGTAAACGATTCTTCCGCCCCCATGGGACGATACCTGTCGCAACTTTGCGGGCAGGAATGGTACCGTACGTTCCTGAACTCACATCTCAAGCATGATCTTCTCAGTACCAGTCCGCACGAAGTGCGGGTCCAAAAACCAGGCTGGTCGTACGTCTATTGCTAATATCTTCGTTTTTACTCTGGTGCTGGCTCCAACCGGAGGTTATGTTGCGACTGAGCAAAGTTTCAAAGAAATGATTGTATTCAGGAGGTTGTCCACTAACACCGCCTAGCATAAACCTATGCGTACTTATCGCCAGAACTCACCCTCTGAAAGGAAATGTGATGAAGTTCTAAACGTCTTCCGACCGACCCCATAAGGTATACAACTAGGCGCTTATTAGACGCGACAGCGCTGGTGT
|
| 7 |
+
>seq_03
|
| 8 |
+
TACTGCCTAGCGGATTAAAAACCATGATCCAGAAGAGGCTTACCACGATGTCGGTCAGGGGTTAGTGGGTTATGGTACTATCGGTGGAAATAGAGTTCTGTCACGTTGACATCCAGGAGCACCGAATTTAGTCACGTTCCTACGGCCTCCACGGTCATTTGCAGATTAGGTGGCGGATCTTAGAATGATGTGGCGTCTCGGTGTTGCTTACCCCCTTCGGTAAGTCCTGCTTATGCATCTTGTTGTGTGAAAACCATAATGCTCAACACTCCAGCCTTGTTACCAGAGGTGCGCCCGGTACACGAGCCGCGCACGTTAACAGCAGGAACACGCAGAAAAACGGGCATTAGCCAAGTCCAGCACTGGAGTACCTTAGATCGGTAACTGCTCGCACGGGCGACATGATGATGATTAGCTCGCTGAGAAGTGAACTTCATGTACAGACTTTGCCAGATAGTTTTGAGTCTTTAAGCTTTACCTACGGCGCGTTCATTTTTAGGGCGTCGCGACCGCGTGCGCCCCTCCGCGGAACTCTCGACGCAGATGAGTTATACCTATTCACGTATAGGTATTAAAACAGGTAGAGCAGGTCAACCAGATCCTAATTAGGTGTTGCTCGGCTAGAGACCTTTAGAAATCTCGATCACAAGGCTGTGTCGGCACAAACACTTAGAAACCTCCTCACCACTGTGATCACGTGTACAAGCTACCTCACGTGTTTTCTACGTGGCTACAACCACATTTCTGATAAGGATGTCCGTCCGCTTGCCATGTTGAGCTAAAGATACGAGGCTACTACTGAGTTTGATTTGGTTTGTTGACAGAAGCTACCTGTTCTCCCGCCTCAGCTACTTCCAGCAGCTAGTCGTAGATAGGATAAAGATTCTGGAGTAATGGCCATTGTCCTAACCCACCATCAATCCTTAATACAGGTAGTTCAGCGTATTATAGCGAACCCAGAACGGACGGCTTCCGTAAAGACTGGTCACACTAATACACC
|
| 9 |
+
>seq_04
|
| 10 |
+
TGCCTAAGGGAGAACGGAACATGTCGGTTGCGTGGCCGCCGTAATTAGCGGTGTCCGCGGATAAAAGATGCGGCCATTGGTTATTGTCAAAAATCCACCCGTTTTCTTTTCTATAACATCAGACAAGATGGGTTTGTCCGCATCGCGTAATGTCGTTCGAAAACTATGTGTACCACACGATTTTACCTCTCTCGTGGAACGTTGATCTAGGTAGATTGCGCGCTCATTTTCAATGTTACCCACCTGGCGGCCCATCTGAGTAGTTGCCTGAGGTACGTCTGCACCTCACGGCAGCGCAATACGCGCTCAAGTAGGAGTCCCCCACCCGGCCGCGCGACCGGGATACTGCCCAGTGTACGGTCCTATTTTCTCAATATTCCTAACCGCTGATGTGATTTGCAGAATCAAGGAGCATTAGTATTACCCATCCCCACAGAACCCCCAGCACATCCATAAAGGTTACAAACCACCCAGCTCGCATTTGGATGACGTTATCGCCGCACCGGCTCTTCCTTGTGTTCACGCATTAGATGGCGAATTCGAATTTTTCGGCCCGCCAGACCTGAAAATTACAGATCCGGTGTACACGTTTCGATCATCTTATTTCATGCACTGGACGGTTGTTACGTTACGTCTGTACTGGGACTTTTTACTTATGAGTATCAGTTTCTAGTGCGCGAGACGACCTGCATAGAACAATGGGGCTGTGGTGCAGATTGCAACTACCTGCACTCGCACCGCTATTCCTTGACACAACAGTTAACAGATGGTCCGACACTGACATGGGTGGGATAAGGGCATCCTTTGAAGTGAGAATTAAGGCCGGTTTTGAAAATACCAAGAGCGGACAAGCGTCATCCATGGGACGACCTGTAACAACCACTTTCTCGACGACGCAGATCCTTGTATTCCTCTACTTAGACCTAGGTGTGAAGAAACTGAGACGACTTCTTGCGGCCGTTGTAGTAATTTCGCCAGCTACCCTCCTCTGGATAAAATC
|
| 11 |
+
>seq_05
|
| 12 |
+
CCGATCCGACGGCAGTCGAGTTGCGGCTTATTCTCGTTGACCGTGGTGCTCCTTAAGCATGTGTTATAGCTGGTAAGATGAGTACATCTTCGTTCTACGTTTTATCCCTCAGTCGTTATCGGCGGAAGAGAAAGTAAACGTAGATCGGGCCGCGGAGATCACAGCTACCCGGCATTCAGCACGGAGGTAAGCTGGGATCGAGCAAAACCAGCGTCATTCAGGATTCTAACAAACTAGAAGACTATTGCATCCAGCGTAGCGATCGCAGTTAATCCCAGCTCGCCGACGGTAGTGAAGTTCGTGCCGCACATTTTCGTATTTTACGATAATCGAGTACCGTCCTCTAATTCTGTAAAGAATGGCGACAGGACTCTGCAGTCCCCCCCACTCTTAACTAGTGCGATAGGGGGTAAGTAGGGACCGTATTTTGCCACTTAAGGTCACTATAGAAGACGCTTGCAATAGAAAAGCGTGGGGCCTGTACCAGCTAGGCCCGCTCTTGAGGACCGTCAGAACCATACTTAATAGTTACAAGGTCGGCTACAGAAGATCGCTAAGCCGATTGACCGCAATGACCCAGCGCCGAGAGATCTTCTCTGAGGGCCGTTAGGAATAGTAATACTAACGCACTTATCGAAAAAATACTCGCGCTCCAGGCATTTATGGGATTAAGGGACCCGTCCGTTCTCTACATCCCCGCGCCTTAAAGGGTGACCGACAACTCTAAGATGATGTTAAGCGAAGGTATAAAGGATCCGGGATGCTGTTAGTGTGTAGCCCTATAAATGACCGCGAGCCATAGTACTTCGTGTGAGGACTCACGTAAAACCACGTAGTTACTGCACGAGTATCAGGAGCTCTTAGATTTTAGAACTACGCCCACTAGTGAGTATTTGGCCAACGAAGGATGCGTCTGCTTCGAGGGGAAAGGCGTCGCAGCTCCTAACACTGCCTCACGCCCGAACGGGGTAGCTAGTACCGACACTACCTGAAGTTAGTG
|
| 13 |
+
>seq_06
|
| 14 |
+
TAAGCAAGGTGGAGAAGCCTGTTACACTCACAGGTTACGATGGACAGTCGTCCGTGCTCCTTCTAAGTTGACATGAGGTGCCCGCATAGGGTGATTAATCTCTCTGCGATACTATGTGGATGCTATATGTTCTGTATATAGGCGGTCGTAGTACCAACGGATCTAATGAACAAGCTCATGCTAAGTCACACAGACGGCTACAAAAAGGTTGGACAATGTGGCATTTAGAACAACGAAATCACCGGCGATTGCCTGTAAGCTAGTCAACATAACTCTAATGTCTGCTGGGATAGGACCATCGGGTCGTTCAAATATGGATTTTAAAACCTTGGCGAGTGTTAGGGAATCCCCTACGAATACGCTCGTCTAGCTGCATACTCACGCGGGGAGCGCTGTCGAGTGCAGATTCTTTCACGGTAACACGACGCGCTCGTATGCCCTGCGGTAGTTACGCTCAACTGGGTTTACGGGAATTTGGCGTTCTATCGTCTCAGTCGCTACGGTGGGCAAACCATCAGCGGATCAGCTCCCGTCCAGTATACCGCGGCCGAAACCGAGGTAACTCCGGGTCGATTTTGGCCATGCCCCACTCCAGACTACGGGTGATGGGATTTTAGGACGGGTACTCGTTACCGAACAGCAGTAGACCAAACCAGTTGTGTTGGTTTTGGGGTCCGTAATTACCACCATCCCTCGGTGTCCAATCCTAGTGATCCATTTCCACAGCAGCACAGGTACCAAGACAGGAACCTGTCTCCTTGAGATCGTGACAAGACGCTAGTAAATCGGACCAAATGCAGATCCAATCTACGAGCCCCGATTACAGGCACTAGACCCAGGCAAGACGGTTCAGGCTGCTGAGAACACGCAAGAGAGGGAGATAAATTTCACCTGGATACAAATATCGCTTGCAATTCTTTCGTGCACCAATTGTGGATACCTACACCACTGCAATGCGCCCTCTTTAGAAACTCACGAGTGCACATGAATCTATGTACAC
|
| 15 |
+
>seq_07
|
| 16 |
+
ATCAAATTAGCTTTATGTAGGGATATTCGCTACTGAGCATCTCGAACAAGAGCATGAGGAAACATGCGCCGAGACGCATCCCACCTATATAGGGAAGCACCCTCTTTCGGCTGGAGGAAACTAAAGGGGTTCGGCTTTGCGCTGGATTTTACCCCATTGTCCGACTACCCGGCGGTAGAATCATAATTATGCGTCCTTACGCTTGTTCGTTAATCCTTTTCAAGGGAGTATAAAAGATTCAATCACTAATCGTGTGCAACTATTGCTCCAGCGGAAAGGTTACGTAAGAGGTATCACAAGGTTTGTAGAGCAGACCAACTGTCATAAAGAGCAGTTTCAGATATATCTGCGCAATGCGTTTGCCCTACACGGAATTTATGAATTGGTCGAACCAAAGGGGAAGAGATTGACTATTCGTCTCCACGAGCTGGTGGGGGTTCTAATTACTGAATAGGTCGGTCAGATTGCTGAACCCGTCAGTGCACTTGACGGGAACATTGTGCGACCGCTTTTCCGCTCTGTGAAGAGCGATCATGGGGTCCGACTCGTTCTGGAACAGAGTTCTGGAGCGATACCTAGTGTCACAGGTAATTTTGGTTCGCCGCTTGTGGTCCTGCCTCCTGCCTGGAAGCCCACTTGGAGCGAATATCAATGGAGCGGTTCAAGCCGTCAAGGACACTCCGTCATCTGGCAGGCTTCCATCCAAGTTAATCCCGGCAGTACTCGTGGGCTACGTGAGTGGTTTTTTAGACGGATGTAGGATGCTAAGTGATCAACTAGCCTGAAATTTATGACAACGCACGTACCTTTGCCCTGTATTGAGAGCTTACTCGTCGAGCAATGTCGGCGGATTACTGGTTTGATGCGCGCCGGGTCTAGCACTAATCTGTGCATCGAAACAGCCCACCCGCAAAACATGTTGCCTCCTCGCTCTGTCGAAGTGCCGTGCACGACCGCAGGTTTAGGGGCCGCCAGGGCGAGACCCATCAGCTCGACACGG
|
| 17 |
+
>seq_08
|
| 18 |
+
GTACACGTATACGTTCGCGGTTGTAAGTTAACCTGAAGTAATAACCCGATGGATTAAGCACCTCATGAAGCCGAAGGGTATCAGGGCGTTAGCAGCGACTCAAGGTCCCTTACGTATGGGGATAGCCCGGTTTTGAAGCGGAGGTGCCTGGAACGAATTTTCCTTGAGCTAGTGATTATGTAGAACACCACTCTTATCCTGTGAAGTCGTTGTTACCTAGGATCTTGCACGGCTCGTCCGCCCTGATAATCTTGGCCCGGAGATTTGTAAAACCTTTTGTAGCTAACATATTGCTCCACCTGGCGTAACATGGTCCTATGGGGGTCAATGCCCGGCGACACCCACAGCAAGATTGATAGCCTCACATATTGCCAAAGGCCCTCTCGTACGTGTCAGGGGCTAAACCTCGTGGCTGTGTTACTTTCAGGTGTATGAATGGGTACTTTCTCTATGAAGATGCAGTCGGTGCATCTATCAGCTGGTCATACCTTAAAATTTTCGTGTACTTCGAGTCCCAAGGTACTGGAAGAAATGTAGACATTAATGTGGACTCCAGTGTGAAACGGCAAGTATACACCTGGGTTGTGAAGCGAAATAAAACAAGCCATTGCACCTGCTACAAAAACTATGTTGCAAGAGAGATCTGATCCGGAGTCCATGTATCGTTCGAGTCACAACCGCCAGGCTCTGAGTAAATCCACAAAATATAGGCGATATTGCCGGTGGCATGCTTCACAGCCATATGCCTCGCCTCTAACGAGCAGTCTGTGGACTCGCCTCAGCGTCGCGTGGGGCCCGACGACCAGGGCCCGTCAACAACAATCGCACTGCTCCTTTCTATCTCGACCAGCTGATTGTGAGCACGCGTGGGAATGCCACATAGAGTTCACACTTCTAGGCACTGTCTAACACTGGTTCCGCACAGCTGTGGAGTCTGAACTGGCAACTCCGTCTGAAAAGATAAGCTTTTCGTCCATGCTATAGTCCCTGGTAGTAAGAG
|
| 19 |
+
>seq_09
|
| 20 |
+
TGCCGTTTGTTACGCGGGGCGGGATCGAAAATTCGATTGTTCGATGAACTCCTACTGATCATAGCTGAGTTATGCTTTTTGTCAGACCCCTAATACTTCCTGCACCTACGAATGGATTGTGCAAAGAGAGACTCAATTTACCCGCTGGCTTCACTGAATTGTAAGATTACTCGTGAATCATGCACCGTTGATACGAATGCCACGACCACCATTGAAGTATTCCAAGCCGTACCGGCGCTGGGCCATACGGCAATCTAGTAGCCCCTCAGCGAGAGGTGATGTCTCAAACTCAAGAAGTCGGCTATGCACTTCCGCTACGAAAAACCCGGTTGCACCTTGGTGGCTCAGGGGCATACGCAAGGCAGTGTGGGTGGACCCAGGCCCGCCCCCAGGATAACGAAAGACACCCGTTCAGTAATGAAGCTAGCGAAGTGAGTTCTCCGGCTGGCTGGATGCTCCTAACATCGTCACCCAAACTGATCCTCAAGACAATTTGAGGTAAGACCAATGGTTTGTAGGCACGGGGGTAGACAGCGTCGTGACTACTGTTCGGCAGGCTTATAGGAGGACTTTCTGTGTGCAACCTGGCACGTGGCAGTCCAAAAACATTACGATAGGCGAGTCCAACTGCGACAATTCCTCAACATACACATGGATCCAATGGATTGGAATTAAACATCGCCGTGACGATAACGACGCCAAATGGCCAGGGTACGCAAATTAAGATTTCTCACCTGATTGGGGCAGGGCCCAGAGCGGAGATCTAACTCGGGGGTGGGTCGTTGGTTCTCGCTAGGCTAAAGTTCAAGGGGTAGCGTGTAGACTAGGCGGCTCTGCAACACATGCATAGAGCACCAGGCTCGCACCGGAACTGGTCGACACTGGGCTATCGTCCATTTGTACTGTTGAAAACCCAAAACTCATCGCCTGTGGTAACGTTCGTAAATAGTGGCTAGCGATGGGGACTCGCCATCGTAACCCCAGGGCACTATATATGTGA
|
| 21 |
+
>seq_10
|
| 22 |
+
TGGAGGTGCGATAAGAAACTTCGGCCCCGTGGAGCCCAATGTTGGGCGTAGTGTAGTTTTCTTTGCCCCCAGGAATCAATCAGATCCTCGTCAATTCGTCGCGACCCTACAGAGCTTACTCGACCAAATCCCGTTCATGATTGAAGTGGTCCCGGCAGCGTCTAGACCCAGAGCTTTACGTCCTGAATCTAGTGCGGACAAACCTAAGGGATCCCACGTAGTCAAGCGTAACCGCGCATAGTAGGACCTGCGCGAGAGCTTGTTGGTAGCAGCTTCTAACCGATACAGTACGGTAGGGCCCGTTGGAACCCGACTTACGTGCTTCTGACGCCAAGGATTAACGGGGTTGATGCAAGATCAAGTAAAAATGGGCGAGTCCGACCATAACTCTTAAAGCATTTTAGCACCCGCTACCTATGTGTACTGCTGCTTTGATATAAGCTATCGGGGGGTACGACAGTTTGGAAGAACCGCATGTCCCCCACCCTGTGGTTCTATAGGAGACATATGCGCAGAGAGTCAACACTCCTCGTTGTCAGATAACAAAGAGGTCAGCCATATCATAATGTAACACAATTGGAAAAGGCCTATTCCCTCATTACGGGAGAGTCGGACCGTGCGAGTCATGTGGTTGACTTATCAAGCCAACAAGATAGGTGGTACCCGTCCTGCTTGAAATTTTTCCGCGGCCCCGGGGGATTGTGTCTCTTGGCACTCGTATTACACTCGGGTAGACCAATATATGAATCTCCCACTCTTGCGCATCATGTTCTATAAGTCCCTTGCTGAACATCGTCAGACGCCGGATCCGCTGCGGCTAACGTGTCTAATATGAGCTTAGCAAGGCCCGGCTCCTATGGAACATAAAAATGGCGTCGAGTTGAGTGTCACTAGTAATGCGCAAGCTCCTGTTTTGCATGACTTTGACGAATCGAGTCAGGAGTTTGTGCATAACTCAGCAGTGTGAACTGCTCGATAAGTATGCCGGCAGAGAAAGTGC
|
| 23 |
+
>seq_11
|
| 24 |
+
TTTCGCCATTGAAAGACTATACGCCCCTCAGGGAGCGAATTTTGCACTGGGGGCGACCCCCATCGTTCGTCGTTCATTATCACACCTTCTGGAATGCAGTACATGGGATCCCATATTACTCACTCATCCTCAATTGACCCGAAACCTGCATACTAGCCATATGGTACCATAGGTAAGTAGTAATCACGGACTTGGTGGCCGACAAAAACGTGAAAAACATTCGTATTGTCTCGATCCCGAGGTTGGGTATACTCTGTGTGGGACTGTATAACGGTACATGTTACTGTCGAGTCGCTTGCAATCCCTATTACCCTATATTGGGGCATCTGCCTCTCCTCGCCAGTGGCCTCCACAAACTGAAGGAGGACGCCTGGCCCGTGACGTTTACTCCTCTCTACATCCCGGGCCTATAAATATAGAGTCTAAACCGATCCGTCGGGAGCTAGAGGAGCAATCGTCTACATTGAATTCGTGGCGAACGGAACGATCAATGCCAAGATCGTGACGAAACCCTGTGGTTGACATCGGGGGCCGGTTCTAGGCCGAAAGAGGCATGTAAGACGAGCAGTTTGTGATACCAATGGTCTCTGTCCAAAGGGCGAGAAGTAACCAATCGTGAGTGCATACATTTGATCGCATTGCCCGATCCCCGCTGCGTGTGACTTAGCCGTGTACCAATCCCGCGGCTAGCGTCGTGATTGCGCAGATGGGAGGACCCTGGAACGTAGCTAGGCGCTGAGGTTCCGTTCCTCCTGGAGACTCACGGAGGATTCATCCAAGTAAGTTTATGACAGCGATGGGTCAAACCTGAGAGGAAGTACCATGAATGCTGGGAAAGTAATACCCCACCTGCCATTGGATCGTTCTCGGCGTAACTTCACGATAATCTAGCGTAGGTGGGAAAGTGCCGGGCGCGCAATCATTCCGACCATGCGAACAAGGCTATGCCAGAGGATGCCATGAACAGGGTTAACAAGCTAAGCACTACTAATCCATTCGG
|
| 25 |
+
>seq_12
|
| 26 |
+
AGGTGTGGACATTCAGCCGCTCATCGCGTCATTATTCAGCTGCCTCCAGCAAGCTGGTACAACTCATACGCCGCGAATTGCACCCAATGAGTCTAATTTTTTGCCGCTCGTCATCAAGATCGTGTTATAAACATTGATCGTTAGCGTGCAAATGTTTACTCTCTTGGTTGTTACACGGCTTGTCAATTATCCGCACACATGCCTTATCCTTTATTAAGGCTGGATGGAGCGTGGCTGATCCCACCTTCGATTACGGACGAGATCGCCCTCCCAATTACCTCAGCATCAAATTCTCGTCAGGCCATTGTACCCGCCGTTGGTTCTTCAGGCTCAGATGAGTCACTAAGACGTAGCCATAAATTGGTCTCAGCATACAAGCGATGGACACGCTGCCGAAGACTTAGACTATCGCAGAATGCTGGGCGCATGTTTCATCAAGGACTTAATTGGTAATTTACTTTTTGTACACACTTCGTAGAAGGAGTCCCGAACCGCTGAGAATACCCCATTAGAAGGTCACCGTAATGCCATGAGAACTAGGACGTCTGTGATGTGCACAAAACGTCCAATTTGATGGAAATACCCGAGGGGTATGATTCCCTGTCTACGTTGTTTACATCAGCTTCATCCTGATTCTTTACTCTCTAATCATGGACGGCTATAGAGGGGAGACACTCGGGACCAGAATAAGGCTAGCCAAAATTAGCCTCTTCTAGTTCTACCTTGATTGACTTCGCCGGTACTCAGGAGACGTAGCATGATTCAGTTCTTTAGTGACCTGCGTATACCCTTATCGGCCCATGACAAGACGGCTGTTAACCTTAAAACGGTGAGCCATATACATGTTTGGGGAGAAAATTTCCCGCAACGAAACTCTCGCGTAGAGCGTCTTTCGAAGTACACCTGTATTACATGCCGCTACGCAGCATCTTAAAAGAATCTCATGGCGGTCGACCTCATTCACGAAGCGACGCCGCTGTAATCCGTAGCGAGCGATTAA
|
| 27 |
+
>seq_13
|
| 28 |
+
GTGGGGTCGTTCCGGTAATACTGCGCCGACACATGAACTTCTTTCTAACAGGTCGCCCGTGTGAAACCCTGCAGATGGCACTGTGGGTAACCTTCATAGTTGTATACATAACTGTATCAAACCTGCTGGCTTCAAGGAGAGCTGCTAATCCCCAATGTTCTCGAGAGCAAAATGCAAGCGCATTGCTCTTAGCGCTGTCCTTAGGCCATGTCAGACCTGGCTTATCGTTGATTCCACATTCCCCCTACATAGTATATGATCGGGAATTCGACGTTGAGTGAGTGTCGTGGTCCCATGATCCCTATTTCCGGCATCAATTTGCTCAATTCACCATCCCGTAGTTCAGGTGAGCAGTTTACGGTATCCTTTACATTTACTCTTCGTTCCCTTAAGACTTCCATACAAGTGTCGGCAGCTGTACGGTATGAAATTCGGTTACGGCGTGGTCAGAGTCCTGGTACGTCTCTTCTAACGTTCTGGAGCCGTGGGAAACCAAAGTATCCCTGACCTTGTCCATATACACACGAACAATACTGAGATTTACGACGGACGAGTGATAAACCCCTGTACCGTGGGACGGTCCTATTGAGAACGACATTTATCCATTAATATGCAATTTGGGCCTCAACGGCGTCTCTCTAAGAACCTATGACGCACCCTGGGGCCCGGCCTCTGTGAGCACCGAATGTACCGGACCAACGCTGTGCCACCTTGGAAGTCCTCTCGAAAACTGTTGTGAGTCGAGTAGGGCGTAAGGGTCTTAGTAGGACGCCTCCTACCACGCGCCACGAGAGCGGACGATTAGCTAGACTCGGAAAAATCCGCCTTACAGTTAGGTATTATCGGGTAGACAGGATTAGTCCGGCAGTTTGACAGCCACATTCTTATTCGTACCTTAAGAAGCATCCAGTACGTCACCCGCTAGTTACCTCGTAGGTGCGTTGTTCATTTAAGCGTGCGAAGTGGAATCCTGGCATAGCTGATGAGCGTATCACCTAGC
|
| 29 |
+
>seq_14
|
| 30 |
+
ACCCTAACAGCGAGACAACCAAAACAGCATGTTTCTCCTCAAGTTTGACAACCATCGGCTTAGATAGCTGGTGGTCCAGTCGAGGATCTGTGACATCCGAGGTCCGGTCCCCCGGCATGAGTAGGTAACAGAAATTTCAGGGTAACCGTTATCACCCGTCTGCGATTATTTGCGTGTTTGTATGGCTCAGGCGAACAGCCATTAGGTCGCCAAGATATGTCTCGTAGACGTCCTTGGATGTCGTTGAATGTTGTACATTAGGATGGTACCTACCACACGATAGCATCACTGAAAGCCAAGCTACGCGGTGGTCAGCATCACGTCTGTGGACTACCATACCCGCTGCATCCTCACTCATTATCCACTCATGGATACGAGCAGAATACGTATCTACATAGCATGCATTAACTGAAGATCGTGAACTCACTGAATGCTGTTTGCGCAGAGAGTGTTGTAGGTAGACTAAGGTTCTTCCTTGAGCCTTGAACTGCCGATAGGCTGCGACTCCGACGAATGGGACGGTCTGGGAGTCAACTCACGATATGCCGGGAAAGTGCGGGCAAGCGACTCAGTTCACGCAGAAAGGCGTACTAATACCAGCCGAACCTCGCCCCTTACGTATGACTTCAAACTCAAGCTGCTGTATCTTAGTACGTGCGCCTCAGGGACACGGTTTTGTGAAAAAATATGGTAGGAGAAGGTTCTTATGCGGATGTTCAATTGTGCCAGTGCGATTTTCACCCACATAACATTCCGACCATGCGTGAGGGGAAATGTATACGTTTTTTTACCTGGCCCCACTTACTAGCTGGCCAAACTGGATCAATACTAAGCGCGGACGTCCACGACAAAAGAAAGCGGAACGCGTTATTTTACGTCTCGCGATTCAGGGCATCAATGGTAATCAACTGCACGGATGTGTTCTTTCATACATATTACCATAAACTCCAAATGAATATATAACATAACTTTCGCTGGGCCATGAACGTTTAGGATGCCA
|
| 31 |
+
>seq_15
|
| 32 |
+
GTTGACACGTGATTGTAGGATCCTGATTCTGTTGGTAATTGATAGATTTCGATCGCATGAAGGTCAAGTGGGGTTCGGGCGATCTCGTACAATCGTAAAGACGTCGGAGCTACGTCGCTAGGGGAATCATTACGCTTGTCTCCGGTTATCTGTCCACTCCTTGAGCGGGCTAAAACGCAACCCGTATACGTAAATACCTCTCTTAAGGAGATGGAGTACTTTAGCTACCTAGGTCGATGGGTCACGGGTGCGTATTAGGTCTCCATGCATGGCCGTGCATCCGCGTTGTGGGAAACTATGCGAGCGTTTGTCGAGTGGAGGATATTGCAGACTGTGGCATCTGGAGTGGACACAACACGCGCATAAGGGTGGGGAAGACGTGGTGGCATAAGCTCAGGTGATTAGAACTACGGATTAGTACTCGAGGTTGATAAGCCTCGGAGCTGCGGCGTTTTTCTCGGAACTACGTACACACCGAGTCGCATACTCTATGACAGTCTGAGACACCAAGCTTCTCTCTGTCAAGTTGACGTCAAAAGCGATCGGCAATCTTTCATGAGATCAGAGTTTGTTGATTTGAACGAGATAAAAACACTGGTATTCTAAGGCAGGGGTCTTGTGTGTGTCGTCGAGCAACATTACTTCGCCATTCGGCAAAGGGCGGCGATCGTCAGCTACGAGTGGGCATGTCTAATGCCGCATAGCATCCCGCGTTTCCCGGACGTGGCTGATCATCCAGCACTGGCTGAGTTACGCTGATAACTAGTATTTAGATCGTCACGAACTTAGGCGCAAGAACTTTCGCGCAGGCTACGCCGCTATGCTTCAGCCAGAAGGGTCGGGCTGTTAGAGGACCTGGACGTATGCGTTGCTTAAGGGCGTCGGCTATCGTTAGTTGGAGAGCCAGCAAACTGCATGCAAAAGCTGATACCTAAATGCTACTCTATATGGCTCTTATTGTGTGCGTATAGGGATTGTGGTAAGTCAAGGGATACCCTAG
|
| 33 |
+
>seq_16
|
| 34 |
+
CCCCTTGAGGAAGGGCGACGTAAGAATGTGCCGAAGGCCGTAAAAATAAATAACTTTCTATACCCAAAACCCCCCCCGCGCCTTAAGTGATGTGAGGCGCCTGAGCGCCCTGAACTGGTGATGGATAAGATCCCTACCTTCACGAGCAATGTGCGAACAGTCACTGTTATCCAACCGTCAGAGGTTGAGGAAGCCGGAAACCACTACGAACAGGTAAACCTCCTACCGACCGGTTTGCTCCGGCGTTTTGTACTACAGAGAGTGCGACTCGTTCGCGCGTTTGTAGACTAGGGTGCTTCAGCTTTTTTCAGTAGCTCCATGACCGGGAACGGATCTCTTCCGAACCAGGACGGACTTTCTTACTGTTCGCCTCAGATCTTCGTAACTTGGGTGATGGAAGCAGCTCACGACTCATGCTGGATGGAAGCGTAAACCACTGACGCATACCCCCTACATGTGCTGGACTTTTTTCCACTAATTAAATTTGCTAGCAGGTTCCGCCCCGCCAGGAAGAGGCATACTAAGATTTCTTTCAATCGGTTACGATATACAGGGGCCGACATGTGAACCAACCATCTCAGCATCTTTCTGGGTCAAGAATGATACGCCCCCTGTTGCTACACGTAATTGGTCGGAGGATTCTACCACCTGACCGAACAAGGGCGACTGTGCACCCGACTTCATGTGTAAGCATCTTTACTTTGTCTCAGCTGGCACGGCGCGTATGATAAGGAGGACTATTGGAGTGGACGAATAACGCGCCTTCTAGAAAGTCGCCCATTATGTCTTGCAAACACCTGTATCCGAGGCGGCGCTCTGAAATGTCAGGAACCACTGAAAATCTTTCGAAGATCTGGCCCGCACCCCTCTTAGTGCGTAGCGTGACACGTTATACTGTATTTAAGTTATTGGGAGTGTAGCCGTCAGTCGGTCTCAGCTGCTATCACGCTTACAACATGAGAGACGCATATGACACCATAGCGGCACAGGAAGTACTGGC
|
| 35 |
+
>seq_17
|
| 36 |
+
CAGGCCCTTGTTTCACTTTATGCTATTGTCCCGGTCCGGCCGAGGAGAGGTCTTTTCAGTAGAGGTCGAGCTCAAGAAAAGGTCATTGTTATAATAGGAGATTTATCCGACTTCCTTCCTGGCGCCTGAAGATACTCATTATGGTTCTCCAGTACGACTTGCGCATTATATAGAATATCGCTATCTAAAATACTACATGGTACACTGACCGCATATCAGGACACGCCCCAGCTTCAAGCCCGCGCCTTGGGCGCTGAGGTATTGTTAAAGCCGGCCCATTTAGCGCTAGCGAAGCATGAACACGACTCGCCCGAACTCTTGGTATGTGGAAGTTTGTTCTGGTTGGCAATGGCGATTGAGCCCCCATCTAGGTGACCACGAAAACGTACTACCTACAAGCAATGCGACGCTTCGCGTGACACGCTCCCAATTTAACCTGCAGCAGCACTGGAATATTCCAAAGGAATACTGTTCCGGGAACAATTTTTGTATGTAGTCTTGGAAACATACCACACGGGGGAAATCCCGCAGCGCTGCAGGTATTCAAAATCGCCTCAGAACTTGTATTTGCCACAACACCCACATGTACGTTATCCGAAAATGGAAGGAGAATCGGGAGCCCCGTGAAGCACGCATTGCCCCTGAACAGTACCGTTATGACCCCTGTGGATCCGACCTTTGGTGAGGCTCTCTTGTCGCCACTGGGACGGCCAAGGAATTACAGTCAGTTTTTCGCCAGTATACCTTCTGTATAGCACGGTGCGGCCATTCAGATACGACCTACAGAATCGCTGATGACCATACGACACTGAACTGGGTCTAGCATTTATTGGGACGACGTTATGAGGGAAGTATGCCATTAGTTGGTGGCACTCTAAAACCGAGCTGCGGGTGGCTGCAAGACATAGACTTTTAATGCGTCATGAAACCGCCACCTCGCCCATGTCAAGGGCGTGTTGGATCACTCTGGTCCAGCAGCATAAAGGCCTCAAAAGTGCCA
|
| 37 |
+
>seq_18
|
| 38 |
+
CTCCTTGTAGTGGTGGAGAATCTAGCTTTATCGCGAACTTTATAGTTCGGTGACTCTTGAACATAGGTGACAAACCGGTTGGCTAGCTTTTAGTATAACGGTATTTTGAGAACACCCTTCATTTTGCCGGTGTCTACAACCCTAAGCCACAGGGCCGACTTCAAGACGCTAAGGAGCCGCGCTGCCTTGTTGGACGCGCCTCTGGGTAACAAGCCCTGGAATATCCCGAGGGCCTTCGTATGTGGCCTTCTGGGTTCGAGCGCCGTTGTGACGCCCGTAGGGCGGCTCATTACGCACCGGTCCACTGTAGAACGTTGTGTGTTGCTCCTATTCGCTAAGGTCCACACCTTTGGAAAGCGTACAAAATTCCATCACTTAACTTGTGTCGGCATACGGCGGTATAATAGAAAAGCAACACCTTAGTCGCGAGACTTATCCGAGCTAACAGTCCGACAAACCAGGGGTATACATAAGAGGTTAGCTGATAAACTGGGTACAAGCCCATTATTAAGCCACTTCGGGTGAATCTGAAAGACTTGACTGGGAATACTTGATATAACCAGGATTGCTCTGACAGCATGGATTTCCCGTCCTCGGCGCTTATGAAGTGGATCTAGCGGTGATACCAAATAACGTGAGAGGATCGAGCGTCATACGACGCGGTGCTACCTTGTGTGCTTCCGTGAGAGTGCAGCCACCAGGCAATCCTGATTAGAGAACGCTCTATGATGAAACACATAAGATGGTTGTCGCTCTAAGGCCCTACAACCCCGACGTCCGAGTAGTCACGCTTAGGCGCGTGGAAGCGCCAGCCTACTTTTAGGAACAAGACTGTGTAGAAATTTAGTGAAATCGTTCTACGGACGGTTTATACATCTCGTTCGTGCCACAATATTTTTCTGGTTGCGTTTACACGGGTGTCGCGTGCCAAATTGCTAGGGTGCGGGGAGTGCGCGGAGTTCCCCCTACGGCCTGGTGACATCTGCATGTGTCCTGACCG
|
| 39 |
+
>seq_19
|
| 40 |
+
GTTGTTAACGGTGGCTTTTCGCGATGACTAATGGTGCGCCCGAACCTCCTAATCAACCGTTACGGCATAACCACTCTGAGGTCCCCGAATTACTCCCGAAAACATGTTCACAGCGAAGTCTATCCGGGTTTCTACGATTTCTTTTCGCGGTTTGGGGAAATCGGGACATATTGTCACACGAGAGAATAGATGCACCACATCAGTTTAAGTCGACTGTACTACCCGTCTTACGGCATTATCAGTTGACTGTCTCGGCGAGTCACTTAGTAAGGCGTGGCGTTACCAGTATCATACGGAATAGTTACGTTTACTAAGACTTCATACGTTAAAATGGTAATCCAAAGAGACAGCGCGGGCAATTCGTTTGTTTGTCCAGAGGTTGCGGTTTGCAGGGACTTTAGGTTGAAATACTACTTTTGGTTAACGTGTAGCGCTGCTTGATAATAGAGCATGTTCTCAACTGTTGTATATCTCATTTTGAGGACTACTACAAGCAATCTGAGCGTCTAGGCAATTTTCAACCCTAATGTTCTCGACACTTGCTAAAGGTTTGGGCCTCCATATATTACGGAGGTCTTAGGAAACGCGAAGTTGGATTACCATGGCCAGGAATTTTAGGAACATCGAGCATTCCATCAGGAGGTTCTGTGTTGCTCCTGTCTGTGTCTCGATTAGATACAATTTGTAGTGCCACCCGCCTGTAGGGTATTTCGCAACAATCTCCTGAGGGTATGCGGCAAAGGCCGATCACGTGTGTTTTACCCTTATAGGGATCTAAAAGCACATGATATCTGCCCGAGACTCATGTATCGGATAGAATGGAAAAAGATACCTAATTGCCCAGAGCTGTCTTTACCAAGCTCCGTAGAAATTAATAGTTCAAGCAGGTTGGGAAAAGAAATCCGATTGGGTGCGGCTCAGTCGCGAACATGAATGTCAACCCAAAATGAGGAGCAAATTCATGGAGACACGTCCTTAACAGCTACAATGAGCGATCCCA
|
| 41 |
+
>seq_20
|
| 42 |
+
ATGTCCCACTCATTTTCGTTGCGAATCTCCATACGACTGCCAGATAGGCACGAGGTCGCACACAATTGCGGTAACGTCACCGCCTAGCACCCGATTTGCCGAGACGTTCGACACGAGGAGGAGCTCTGCCTCTATGTCTGAGCTGACGGTTGCGTCGTAACTTCGTAAGGGTTTAGGAGGGATAAGTTTTTTCGGACACGGCTTGCGAACGCGTTATTTTGGTGGGTCTGGTCAGGCGAAGCTGACAATCGAGCGAGGATTATCCAATTAGCGGAGAAAATTCGACGCCCATCCAGGTGCTATTTTGAGCGTGCTCTGGCTCTAGTGCAGCAAATGCATTTATCTTGCAATGGCCTATGCTAGTTGGGTGGGCCATAGCGGAGGCTGTTCCATAATCGGAAGTATGCCTCACATTGACAGGATAATTCAGCGTGACAATTTGCAAGCCTGTTGCCTGCCTCGAACCCGACGGTTTTCCCGAGACGTTCAGAGGCGTAATATAGGCGGTACGTCGTATGGTGACCCTAGAAAATGTGGCCCCAAAGCCATACTGTTTACCGTAACACTTGTCAGTGCATGTAACCTCAAGGAGAATTCCTATCCGTGCAGAATCATCGCAACTGGCGCTCAGGGTAGAAGACTACGAGATTTAGACATTAGACCCGAGAGTGCCACAGACCAGATGGTAAGCCTGCAGCTTTTGTTGACGTACACGCTTTATCAGAGCCCCCGCTCTGCTCCGAGGTTACCGTAGGTAACTAAGTCAAGAAAGGGTGCCGGCCCTGCTGGGGTACCAGGAAAATGCCAGAATACTGGTAGTTCATCGCTAAGGTCTTTGCACGCCGAATGTTCCGCAATCGCTCCGATTTAACATAAGCAAAGGGTGCTTGAATAGCCAAATGTCCGTATCTGTATGTCGGCCGGCCAGGCGGTCGTAACTCAGCATCGGTTTTAGCTCTCTCATGTTAAGACTGGTCAATGTAGATTGTGAAGTTACATC
|
| 43 |
+
>seq_21
|
| 44 |
+
CTAGGCGTCACTCTCTTGGCCGATTTCAGCGGCATGGTCCCAATGCTATGGAGAGTATTTCACGGCCGCCGTTTCGCGAAGATCACCGACGTTGTTCACCTTGTGTTAGCAGTGTATCCGCTCCAACTGGTAACCCCTGCCATAGCCTGGCCCATCCCAGACACGGTGAAGCTCGTTCAGATTTTTGTACTGTGTCCGCTTTTATTGAGTACGTGACGAAACCTTAGAACGCTGGGAGGGCAGAGTTCCGAGCATTCAGCGACGCTACCCTAAGTCGGCAATCCCGGTTGTCGATTTTCACTAGGAAGGCGCTACTTTCAAGGCTGGGTTAGAGGCCCCATTAGACAGGGAATAAACGGTTTGTGAGCGTGGCAAAGAAAGACGACGTTTAATTTGGAGGTCGCGGACGGACTTCGGATCTTGTATATGGCGAGCGGCACCAATCTTGTCCGCAGATCCTGTCAGAGAATGGTACGTCCAAGTACCGCTATATATGTCGGCCACTGTATGGTCATTGCCAGTTAGAAGATGTTTAGGGGCCTTCTTATCCCTATCCTAGAGCCTTCTGTAACTCACGTGCAGACCCGAGCGTCGAGCTGATCGTTTGCAGCCTACTTAGTCCAGTGGGATTCATTGTTCACTCAGCCTTTCTGGAGGCGATTAGGTTTCCCTCGGTCCACAAACCAACTGCGGATTCAGAGGCAAGTTAGGCGTTTAAGCAAGGGCCGGTCATTTTGAACTAAGTTGAGTCATGGTTAGTACGCACCAAAAGGAGGTCTGCAAATCACGCACCACACTAGGTGTCAAGCACGTATTATCATGGATTCCCAACCAATATTTCATCTTAGTTAAATATAGAATCGGGACGCGAAGTGCGTAACATAGGAACTGCGGAAGCTTATGTGGCTGGGTGAACTGTCCGAGGACCTCGAGCATGATCACAGGCCGGCTCACCCGTCTTCGCTGTGTGCCACATAATTCGGGTTTGAGATGCCGCTTG
|
| 45 |
+
>seq_22
|
| 46 |
+
GACTCCCTAATATGCCGGATGGTGCGTGGTAAGCTCATTAGTACAACTTGGGTCGTCGTTAATCTTTAGTACCTGGCCGCGGCACTTTCAATACGATGAACTCGCCACGTAAATGAATTGTTGCATGGCCCATGTAGGAGGTTAACGTCAGCCCGTTACGAGGGGCCACTCGCGCAACATCGGGCTAAACGGTTGCACAGTTTGCTTAGACGATGAGGAAGTCACGTGCGTCCGTCCAAGCCTAGCGATTGTATGCAGGAATCTGGTGCACCTTCGCTGATCTCCAGAGCATTGCCTTCTTAGCCTGCGCCCTGTGGGGTAAACAGCAGGCATCTACACGCATCTAACGTAGTCTTCAGCCTTCAGAAGAATGCCGACACAAGCAAATAATGCATAAGTCCAACTGCCGCTGTACACCTCACCGCCAACCATGCAAATCGGAAGTTACTCTGAGAGAGCAAGGTCGAGTCGACATATTGTTGACGTTTATCCCGTATCACGTCTCGGACTCCAGCTGACTTGGTTTCCACATATTTAGGACATCTCGTACGCATCTTTAGCACTGTGGAGTCAGCGTCGAGCACTCTTCCAGGATGTTTTTAAACCACACCATGAATAGATGATTCGGGTTTTCCAGATCGACCCCGCTATCGTCACGCATACTCCGACCACATCATTGGTCACAAAGGGGCAATGACTAAGGCCCTGTCATATGACACATTGAGCGCTATTACGACGTCCCTTGCGTACCGGTGAACGCCCTCGCTACTTTCTCTCCCACGTTTTATTGTAGAGGCGTTACCGCTGGTCGTTTGAACATGGATAGCTACCTTGACGTCTGCTAAGCGGCTTCCCTCGCGGCAGAGCTGTCGCCTCGTGCATTCAATCACCCACCATAGGTGTGCGCATCTCTCCCGGCCGCCACTGAAGTGCTAGACGTCACCTGTCGGCCTTAGCCGTAACTCGCTGACAATTATACGATGCTGCTGGTCGTACAGTA
|
| 47 |
+
>seq_23
|
| 48 |
+
ATTGGTCAGACGTTTCCCCTAGCTCACGAACATGACGGCATTTTGTAACCTGATCGTACGCTCTGACGTTTGACAGGAGGCTGGGCCGAGTCCTTCTTTACGCCTCTATTACCGTGGCGATAGACTGATCATATCAGATCACCGGCTTACCATACAAGACGGCAGATTTATCTCCCAGGTGTCTCTTCTCGATTTAACTATCACTTGCGGGCACACGCGCCTAGGTCTTTGCGCCAGGAACTCGGCGACTTTGAGGCTTGCCATTCGTTATAGTTCGTCCATGATGCAGGAGCCCCAGTGGCCGGCATTAATAGTGACTTTGTCGACGTGACTGGTTAGGGGTTTCTCGCCAGGGCTCAATGTCCAGAGGGTCATAAAGACAACGCTAACGCTCAATCGCGAGGTTACGCGGTAGCCCCAGTAAACTTATGGTCTACGGTGGGTAGGGAGAGTATCTACTGCTCTAGCTAACTAGAGAGACACAGTTCGCAAGTTTGGATGTGGCTTGTGGTCGCAATTTCCCTCATGGACTAGTCAGTCCCGAAGGAGGAATAATGTGAGGCGAACGCCTGACCGATATTCGGGCCTAGCTCGAAACAACAGAATTTTGTAGCACATGCTCTGTTGAGTTTGCGGTCGTGTTGTACGTCCAACGCCTCCGTCAGCAGGATGATCGGATGGCCCATGGGCCCCTTGCTGAGCCAAACAGATTTCTACGATAGGCCACAATGCATATGTACAGGTTGGACCTCTATTGGATGTTTCATAGATCTTGTTGTGGCCTAAACACGCAGCCGTCCGCCCAGACATCTGCTTGAGTACCGTACTTACATAGCGGATAGTAGGGTCTGAGAAAGTGATCTCTCCATACTTCTCATATTCGTTTCATGGTACTACAGAAGTCACACCGCCTGCCCCCACAAAGACTCCTACCTCATAAGTGCCTTTTGAGGACTAGGTCTTCAGGTGTGTCTCATTGCAAAACCATTCTAGGAGTGGG
|
| 49 |
+
>seq_24
|
| 50 |
+
TCAATCCAAGCTAGCAAGAAGCTGATTGTTTGGGGGGATTGTCTTAGCTCAATTTTCGAGGTCCATTGGCGGCGGCTGATGTCTCGCAATCCGGTTGAACCAATACGGGGGACAGACTTGCCTCCCCGCAACCTAACCGAAAGCGGACTCGCGAAACATTAATATCCCGAGGGCCAGTCCACGTAAAGAAGTTATATCACGAACTATCCTAGCCCCTGCTCATTACTGAGACCATAGATTTCGAGGACACCCACGTGTTTAGGACTTCGTTTGGGCAAACAACACGGAACTATTAACTCCCGACATCTGATAAGCAATAGACCATTGGGGATCTTCTACGTGGTGACAGTTCGCTTACGATCCGACGTTTGAGTAGCCGCAGTTCTCATCCCAGTGTTAAAGGACACGTTACGCATGAACTCCGGGGACTCCGCATATAAGCCGGGTACAGAGCTCTCCTTTTAGCGGAACGTACTCTGTAAAGTATCCAGTCCAGGATAACCTGCTCTGCTTCGAGAAAAGGGGGATTAATGCCTGTGGAAACCCGTTCTGCGACACTGACTTATCAAGGTTAGATTTCCCACACCTTCTGTACCACACCGAGTCAAAACCCGGGGCTCCGCCGCGAAGTAAACGTTTAATCTCGTGGCGCTAGATAGCCGTATCCCGCGACACCTCTCTCCTGCCAGTAGGGACGGATATTGCAAATAGTAAGCGGGGGTCAGGGCCCGCCACCCCACCGATTCTTATTCGCAATGGTGCAGTCTGCGTGACGCTGTCCCTGGAGGTTAACGAGTCCAGCGCGTGATTCAACACCGGGGTTCCCCATTTACTATGGTACACTCAAGTCAAGTGGCAACCACCCTGTGGTTTTAATAATTTACTTTCGAACGACCTGGGTGGCAGGTTCGACGAGTTATTGTCAGAGATACTTCTCCAACGCTGTCACGCCTACCTGAACCTCCAATCGATACGACTAGCGCCGGATAGCGCAGTGCGC
|
| 51 |
+
>seq_25
|
| 52 |
+
TAGGACCGATGGCCGAAGCACTTATGGTGGTAAGCGGTAGACGCTAAAGACAGAATTCCTCCGAGTCACATACTTATGAAACAATCCATTTTCTTGCCGATCTGTCGACGAGGCATCTACGCTGTCCACCACGGCATCGCTGTGGTGGAATTCTTAATCTCAAACGAACATTTAACCAACATCCTGGTACCGCTCTTCGAGTTCCTAATTGCCGAGAATCTTTGAGGATAACTTGGGTTCCGCGCATCCAAATCGAGGCGGGAACCATGAGGCGCGTCGGCCCCGTTCCAAATTGATATCGCACCATGTAAATGGCCTATGATTCCTCCACGTCTTATGCCCTGGATACTCCTAACGAGCGACTAGGTGTATCGCGTCGTGAACTAGAATCCTCTCATGGGGGCCACTAGGCTGGAAAAGGCTTCATTGGTGGACTGTCTAGACACACCTGGCTCCATTCTACGTGTAAACTTTTTCAACTGTACTCTGGTTTTGTGCGATTAAACTTACGTGTAGGACGTGAACGAGCGTAGCCGTACTATCATCCCTGCCCTTACAGTTGCGAGGCCGCCCACCCTCACACACGTGTGTGGCATTCATGGCCGTTGTGCATACAAAAGGTCTCAGCCGAGGTCTACATGAATCCTATAAGGTCCACAACCCACCGAGCCGGTCCGAAGCTCACCATGAAAGACCTCCCTATTAAGACCAATCTGTTAGGAACACTGCATTCCCAAGTGTTAGGCGCGGCGTTACCGTAAAAGTCGAGAGCTTACAACATTTAGACTGGGCCTCAATCGGGTGTCTGTTAGGACGATAGACTTCCTTGTAGGACTTAATCTGCACCGTTCTTGCGCTAACCTAGACGTTGAACGAGTAGGGTTCGCGGCTGCCGAGCGTAGATTACCAATCCGCAGAGGTCGCAGTGAGAGTTGTACCTCAGCAAATAGCTGTAACGTTCACGCAGGCCTCAGACTATCCACAAATATAATTACTGGCC
|
| 53 |
+
>seq_26
|
| 54 |
+
CGACCAGCGATAGTACGGCGGACGTAAATCTGCACAGAGAAAAAGTCGTCTTACCGTACCAATGTCCACGGTGGCAGGCGGACGCCATGTTTTCTAGAGGTCCCGCTTATATCTACGTTGCCGTTCCCAGATTGCACTTGGTGCGGTCCATGTATCGCCGTGAGATCGTCGATTGTCTTTGGGTTCTCATAGTGCGGGGGTGCTAGGAGGGTTACTTGTCTTCTCCCGCTTAACACACAATGGGGTCTGTGACCAACCGGTGCACGGAGCATCTCGGAGAATCTTATCGCATGGAGTTATTATTACAGGACCTAGGCTAAGGCAGTATTGATGACCCGGCTCCATGTCGACTCCGTGATCGCTGCCTCGCATTATGCAGTTTCTAAGCTCAAGAATGCGATGTGTTCGTTGAAAAACTTCAAATGCGAGACTCATTCTGAATTCCAGTACAGGCTGGTGCCTCAGATCCGAACGCTTAAAATTAATACTGTACTATCGAGCACGGCAAAAACTCGATAGCTGGTCGGACTACGATTCATTGTCCAACCTTTGATATCACGACATTATGGGTAACCGAACCGAGAAAGAGTCTTCGTCATCTTCAAATGAGGCGAATAATAGAACAGACAGGGCCATTGTACTTCTATTGATTTCACTCGTTCATGTCGAATCTGCCAGTACAGTGTGTGTGGCCGGGAAGTTCGGTACCCGCGGACGCCCGAAAGCCTCCGCGACTATACAGGGGCCATTCTCAGCCGTGCTACGACATCGTAAAGGGCAGTGAGAAGCTGAACTGCATTTAAGATAGCGTCTCTCTCCCATAGAAGATTCCTATAGGTACGAAAAATCCTGTTGATAAGCGACCACTAGCCTGAACGAGCTCAACGCGCCTTGGTCAGTCGCGTTAGCGTACCAGAGACAGGTCGCATCGTTACGCTGCAAATATCTCGCGTAACGGCGGTAGAGCAACTTATAGCACGGGACACAAAAATAAGTATAT
|
| 55 |
+
>seq_27
|
| 56 |
+
AGGAAGAAAGCAGTGGTCTCTCGTTAGCGGTCGCCATTGTGTGCGAAATAATGGGGCGCTCAGTTTCACCAGGACATCTCCCACATGTTGTGCTAAACGCTTTGACGATCCGTCGCTCATGATGACGGCGGCGGCGGGTTGCTGCCTTCACCGTGTGAGTGAGGTAACTCCAGATGCAACCGGTAAATCGAGCTGCTCTAGGAATCTCCCATCACATTTTGGTAATAACAATGTCTTCGTACACCCCCTAAAACCACCTACCCCAGTCAAATCAACTTCATTCAGGGAGGGGAACCCCACGTAACTTCAAAGAGGATTATGGTGCGGGAACAATTGAGAGGGTGGTTCCAATCTGTTCGTCTAGACTACAATGCCGTGATCTTGGAGTTCGGCTACGACGCCGCCCCCTTCTAAGCGGATAGGGAACTTCTTACTATGTCCCTCTGTTATGGACAGTAAGGACATCACAAAATCCCGGCACGTGTACAAGCCCACGCTCGCCGGATAAGATAAATCCGGAACCACCTCGGGCACACCAATGTCAGGTTCCTCCCCAAGAGTCGTTATTAGGGTTGCTATTACGCACGGAGTGACCTACGGTTACCAATTTTGCACTAATGTTTTGCTAATATTTGACAAGAGAGCCTTTTGAGGTTGGCAGGAGAAAAGGCTGGATACACCGCCCGTTTAGTTTTAATTTGCACCAAGTAACGTGTCGTGTATACTACCTATGGACCTGTGCCGGCCTCATTTGTAGCCACAAGCGTCCAAACTAATCTAACGGCCGATCAGTCCGTTGAGTGTTTTGTCCAGACCCCCTCATGGCACTTTTTCTTCGTCAGTTATACTAACGTTATGTTTTTGCTGATATCATGCACCGGTGAGTTTCGCGGAAGTGGTGATGAGTAAAACGAACTCGTGTACGTATAGATTCAAAGTCGTGCCATACGGACCCTTCAGGGTATGGGATAAATAATGGTACCGGGAGGAGGTAGCATAT
|
| 57 |
+
>seq_28
|
| 58 |
+
CAGCCGAGAGGAAGGGAAGAGCGATCCTTAGCTGCGCCCTATAGTGTCGGGACTATTCGAACCCCACGGTGGTTAAGTTGCTAATCCTAAAAGGCGCCGATGACAAGGCAGCTAGCGCTAACGAAAGCGAATCGGCCGGGTGGGCGACCCCCGCGGGCTCATAGATCGAATAACCAGGCTGGGCGCCTCGATAACATTGTTAGTACATCTGAGCTAAGACCCTCTGAGATTTCACACCCTCTGTTATAATCAGAGGGAGGAAGCTCAAACCCACTGCGTGCATTACGCAATGCTTACTGGATGGACCATCTTCCGGCCATAGCACACTCAACCCGCGACTCGACCAGGAAAAGATTATTTGTTCCCTGCGACGTTTATAGGTACCGCCAGATAGGTTGGAAGGCGCTCGGTCCTCACCGTCGACTGGGCACAAACCTGGTTAGAGAATATCACTATGTTAGCCTCCTCCCAGCAGTGCGGAACTTGCCTCCGATCGGCTTTCACTACAGCCGGTTATGCGACCGCCTAGATATCAGGCATTCTAAGGAGATCGGTACATCCTTCTTTGTAACTTCGTGAGGAACTTAGGTTATTTTGACCGGTTGGTATTGTTTCGTAGTTCCTCTAGCGAACCTCTGTCTATTAGTGTGTCCCCACTCGCTACTAAAATTTCAGGGGCACCGCAGAGTACCGGGGATGCGATTAATTGAGCCCGACAAATGCGGCGCAACCCTAGCCCATGCGCAAGAGATGCATCGGATTTAGGAGAGTGAAATTTCCAACTTCAAGAAGCTCCGACCAGTGAATAGCGACCAAATTAGACGACAACACAACAGCGAGTTCTCTGCCCGCTAACGAGGGTTTAGGAGTTCCGACAAGTCTTGATATCTAATCCTGAGTATAGTCCAATAGCAATTGGGCAGTGTATGGGTCTGCTGGGAAATGCACTCTGACAAAGGTCGCCGGGTCGTCGTATGAGCGCTATTAATCACGATAACGG
|
| 59 |
+
>seq_29
|
| 60 |
+
ATATATTGGCCTTATCCCTGCCGCATGAAGCCCACCCACTCACTAGATTAGGTATTTCCCGTCGTGCAAAATGGACACGGAAAATGCTTTGATTACTAGCGTAATAACTGACGTCGCAGAGACTTCGGTAGAGAATATACGACTACCACGATGTCACCTTACAAGTAGATCGTCAAGGCCGGAGGTTCGCGTGCACGCATCTTTAACCCTCCTTGTCAGTCCAGGCCTATGCTTTATACAATATCCCAAATCTGCCTACCATAATCCGAACGTGATCACGGCTGATGTATGCTCTTTGTATAACTAAGACCGGGTGCGAAGCTATATTACAGAGTTGACGGCCCTGATATCATTAACCGTCGCGCTTTTATCGCTACCACGCTGCATACGTCTCTTCTGCGCGTACGAAGTGCAGATCTGTCTAACGATCACACCAAATTGCTCAGCTGTGACGTAATAGTACCCGGGTATCACCGTGAGGGATCAATTAATGTATCAGCTAGCAGTTCAGCGAGAATTTAATGAGATGTGCAAGGGATCGATTTTGGGCTGTTGAGATCCAGTATCTGTTACCGGGCTGTGTGTAGTTTCTTTCCGACAGTAGCAAAAGGCCCCGAGACACCAGTTGGCTCTACGCGGCGCGGCACCCGTGGTGGGTAGCCTTCATGTGGAACAAAGCAGTGTTTCGGCAGCAGCAGCTCGAGACAGGGGTCCAGTCCATGGATTTAAATGTATCCGGCCGCCCAAGGCCGCGAGCTTCTTACCGAACAGTCTGACAGGCAGACTGGGGACCCCGCCAATCAGATGCCCCGGTGACCTATCTCGAATATAGGAGACAGGCGGTACAAAAAAGCACCTTGGCGACTGCATGGCATGCGCGCAACTCAAGTCCACTTCAGAGGAGTTATTGACCAACTCACCATGTAGACGTTAGATAGAATATCTGAGCAGCGTGCTAACCACCGTCTCGTTAGAGAGCAGGCATATATACATATTCCGC
|
| 61 |
+
>seq_30
|
| 62 |
+
GGTTTCCTTCTACGCAAAGTCAGGTTACCCTCGGACGGAAAAGAACACCCGGCCCGGCGTGAATTAGACTGTTACTTCCCCAAATCTTCATGCTCTTAGTTCTAATTCCATTTCTGTGCGAGGACTATACATGCGAAAAACCAACGCTGACCCCTCCCAGCCGGGTAATCTGAAGCTACTTCACGGCGGCTTATCGTTTTGAAAGCCTTTTTACTTCGGGTGGCCGTCCCATATCGCCGCAGCCTAGTAACGCCGCATACGCCGCCAGGTCGATAGAGAATCTGTGGTCCAACCCTAATGGCATAGATAGAATAGAAACCAGCAGCGTTCTGTTTTAGTCAGTTTTCAACGCCGCCCGTGATGTCTTAGGCTCGTCGCCATAGTGAGCATGGAGGGAGAGTCTCGCCTGCCACGGATGTCACCTACGCCATCGCAACTCAAATACCAGTGTCGTGCACTGATTCGCTATGTAACTCACGTCTCTCTCTGGATTTGCCTCCGAGTGGTATTAATGCTCTAGGTACTGTCACGGATCGAAATGCGGCACACGTCAGCAGAAATAAAAAATGCGTTCGTCGAAAGATAATACTAGCCGAGCGTATGGGCCAAACCGTTGGCTCTCACGGTCATTGTAGCAGCTTGGGATACTCGAGAAGTAGCTACTATCTTAGTACGAAGCGTTCCCGCTCGCAGGTGTCAAAACCTGACCAAGGGGCTGACCTACCAGGCATTTGTTAACGCACTGCCCACACATGTTCTATTTCAAATTAATTTGTTGGCGTTGCAGTGTGATGGGTCCGCTACACAACACAGTCTTCCCGATGTACCGTTTAATAGGGAACGTCCTGGTAGTTTAAGCTTGCTTGTAGTCCTGTGCGCATGAGCTAGATGTTCTGATTAATATCATGGAGAAGAGAAGCACTGAACCCAAAATCACCTGATCTAGTTGTAATTGTCTATGTTCTAACGCCGTATGGACTAACCATTCTTCGGCTATCTG
|
| 63 |
+
>seq_31
|
| 64 |
+
CGAGCAGAGGGCGAATTTTCCCGAAGTCATCATCCTTTGTGGACCCATGCACACCAACACTTCTGACCTCCAGAGTTAAATGGGTCCCACAGTCATAAAAAGCCGGACGCATACTATGGTTTCGCGGCTTGTTGGCCTCTTCACTATTACGCGTCGGCGAATCCACCTTCCAGATACTACGTATGCCGAAAGCGGTCACTTTCCGAGGTCGGGACTATAGCCATCGGCCACACCACAACGTTTTCCACTGCCACGCACGGAGCTGACGATCCCACGCGACAAAGGATGTGAACTGAGCGTTAGGTGCGAATCAGGACGCCACAGCCGAGTTCTTTAAAATTTCTCAGCCGTTCGTGCGCCACTATTAGCCGTCCGCACGCTATCACTCGGACTTATTTTTCTGGCTCTAGACACCCACTGCATTGTGAATGCAACTGGAATGATTGCTCGGGAGGAGGTAACGCGGTTTTCACGTGCTCGTGCGGGACCCTTGAGGCTAGGTTGAAATACTTGGATAGTAGTCGAGAAAGAGTAGCGTTAACAGGAGATCTCTTAAGGAACATCAGTATACGCAGGCAATCGTACGACTTTGGCATCGCCAACCATAGGCAAGTAACGATGACGTATGTACGGCGGTCCCCGGCCAACGAGTGTTCCACCGATCGAAGTCTACTCTACACGCATTGGCAAGAGCCCTCGAAAGGGGCTCTTGTTCTTCGCTTTGTAAAATAGTGCCCGGCGTTTATACGCCCGGGCAGCGAGAAGGAGGGACGGGAAGCCCATATCAACCACTATGTCGTCGATGGAAGCTGCGAAGGAATTGAACGGTGCGCGGGTTGCAACGATAGTAGCTTGAACACTTATCAATAGCATTTTAAGCTCCGACTCAGTCCTAACTCGCCAAAGTGACGGCAGCAAGCTCACTAGAAAGATACATAAGAGATAGGTGCACCTCGGTACTTCGGCGAGATGGCTACCGATCGCCAACGATGAAGCATTT
|
| 65 |
+
>seq_32
|
| 66 |
+
CGGTGAATTAGTATAAAAGTAAAAGCCTCTGGAGGAAGTGCTTGCTTGCCTCCGGCTCCAGATCAAGATGTCACAACTCTTCTGCTCTTGGCTGGTTTAATGAGCGTAGCTATACAGTGAGGAGGAACCCCCCTGTACTTTACATGGGCATTAGCGGGTCGTTTGAAAGGCGTGAATTCACGTCCTACTCGACAAATTCGTTCCCATTAGGCCATTGCCGATTGAGGACAATGACGGTTGTCGCCCTAAATCGCACGTATCCAGCGCTTTCTTCAGATGTGGCTTCTGCTACATGTATAATTAGTGGGGATCCGTCGTGAGATTAAAAGCCTACGTTTTACATACCTGTAGGTTAGGCTCGTATCCGCCGGTCATCATCACAAGCAAAAATTGTGCAATGACCTTCTTTCTGTCGCGGCGAACGTACGATAACGACGGTTTTTCACTTGCGTTTCTGGATGCGTCTCTACTCCACAAAAAGCGTGCATTCCCCTGAGGCCGGGAGAGGACAAAAACTTATCGACTGCATAGTTGTGTTATCGAGGAACTCGAGTACTCGCTTAATGCCGGGCGTTACCTCTGTGGAGGGTCAACGTGGATAGCGGCATGTCTATGTCCGTAAAACGTTAAGATCTTAGAGCAGTGATGGAAGCGTGGACGAATGACTAGGTCAAAGATGCTTACCCCCGATTTATGCCGCTGATCTATACAACTACTGCTGCATTGAGGTCCACTTTCCGTCATCACACACAGTCGCTCTGTGACAATAAAAAACACTTTTCGTGAGCAGATTATCCACGTATGAAGATGCTCTTTAGTGGGCAGTGCAAAATCTGTAATCATTGGAATGCTCCACACACACAGGTCTCCCAGTGCATCCGTTTAGCATAGCCCGCTGGGAAGGTATTCCACAGACTTCAAGGTATGATTACAAGAATCAGTAACATGTGGGTCCGACCCTTGAATCGGCCACCATCGGTAACCTATGACCGGTACTCCA
|
| 67 |
+
>seq_33
|
| 68 |
+
GGATGGTGCAATTTATTCTCGTCAATTTTCAAACGTCCTGCTAGTATATATCTGGATAGATTAAACAAATGCCCGTTATCAGTCGCGTCCTCTTCAACAAGAGCCCTTATGGATGCAAGGACCAGCCCTCATTCTTACTGTGTCGTGCAAAGCGAATCATCTGAGTGCCCAATACCGGTCGCGTAGTCCCAATTTTGTGAGCCCCTTCCAAGGGAGCCTCAGTTCAGTCAACTTTGTACGGCACTGAGGTAATGCGGCCCCTCGACCCGGTCCACACGCGGAGAGCGCAAGCATGGACTTGGGGTGGACAAGCCTGCACATTGCAGGCACCCCCGGACTGTGTAACATTGATAGCGGTTCGTAATTCCAGAGTCGTCACTCCCGGCCTGTCAAATCGAAATTACCTATCTGCTGTAGCGCGGATAATCAAAGGGGGAGGCGCACTGGCAGAGTGCGCTAGTACGTCAGCAGGGCATATGGTTTTCGCCTAGACCGCTAAACGACCGATATACAGGTATGCCATTTTGTAGTATATAAGATCCCGCACAACGATTCCTAGTCTTTTTAAAGAACTTCCCGGCACGCAATTTGGAAGGATGAGATTCGTTATGCCACTAGACGTTCATACGCATGGGCTGCGCCGGATTACCAACTCGCAAGATTCTTGCAATGGAGTCAGCGTGATGATAAGGGGAGGGGTGATGCATGGACAGCCCTACAAGCCCATTAGGGCATGATGGACGATGAAGCCGCCCTTAATTGTAAGTGAAACCTTCGCCTCCTTGGCGCCAAGCGACAAATCTAGATCCTCAACGGCTTTGGACACGGAAATGAACCTTGCAATATTCGGTCTCGGTGCCAATGCAGTCTCATTCAAAAGACTTTCCGGCATTACTCGTCAAGACCCTCCCCATAGCGAACGGTAAGATCCCGGAATACGAAGAAGATTATGGAACCGATAGACTGTTCCTGGAAAGATTCCTGGTGGGTGGGGTAGAAT
|
| 69 |
+
>seq_34
|
| 70 |
+
CGGGACCGCCCTCTTTTCAACTAGGGCGAGACAGTAACAAATTTTGGCACGGTTAGTTGTCACGTGACCGTTGAGTCCCTTTGCTTGCTTATATCCTCGACGCAAGGCCCAATGCTCTAGCGAGTCAATCTCGAGAACATGCGTGGCGGGTTAGAGGTAAGAAGGCGCACAACGTCCAAACGCGGCTGTGAGCCACTACACGCAGTAGCATCGATCATATTGTCGTTCCCTCGTGTTCCCAGTGCCTCAACGGTCACTAAGAAACACACCAGTTCGCCTAGCGACGGTCGTTGATGCAATGATCGTGCTTGGCTGCTCTGCAAGATGCACTAGAGTCAACTAATCGGTCCTTAGCAGTCAAAAAAGGACGACTCGAAACGATCATCCCCTTCGAGGTAGCTGAGACTGTGTCTACCCGAGAGGCCGAACCCATTACGTAAGAAAGTCTCTGCTTTCGTAATAGGAGATAACCGTAGCAGCGAGTTAAACTTCTCCAATACTTAGTATACACCCGGAACAGCGGTAGACTCTCATAACGAACTTCCCAATCTGGAAACCTCGAAGGGCCGCTTACGTATAGACGCCATAGATGCAACCCGGAAGGCGAACGGGATGGTAAAGCCTGGAGCGGATTGCGATATGCGCCGAATTACTTGGAGCGCCGACTGACTCCACTCGCATTAGTCAACATTAGTTCTGGTATTTTCCGTAGTCGAACCTTTACTGGAGCGGACGTTCGACTGCGGAATATCCTGTGAAACCAGACTTGAGTAAAAGTAAGACAAAATACCCCATGCAGTAACGGGTACGTAACGATTCCTACGTGTTTGGCCTTCAAGCACCAGGTATCAGCGGAGAGCTTGGCGCCAGGCCCTCACTCGGGGAGCCGGCTGTATCAGAGATGATAAGTTTCGGTAGGCATATCGGTGGCTGTTTTCGGAAAAAGCTTCAGAGGAAGCTCCGGGATTTAAGGTGGGAGGATCACACATAAGACGCAACA
|
| 71 |
+
>seq_35
|
| 72 |
+
CCTACGGAAGGCGTTAGGCTACTGGATTACGCCTCACAGCAGTCGACTAGTTAATTAGGCCGATAACCTGGCCGCCTAACGGTATTCCTTTACCCTGCCGATTTCATATCCAGCCGCAAGCCTATCATGGGTACTGGATAGCTGCACTGTCATATTTTTGACAATCGTGGGCCTGTGCGGGTTAACAACGACGCTCTGTTGGAGATCTTCCTCTTCTCCGGCTATCGATCGGTTTCTCCGACTTGCGACATCATCATGTTAAGTGTAATCTCACCGTCGTTAGCCGTCACCGTAAAGCGGAATGAATCCGGTAGAAGGGTCAGATTTTCTCTCAGTTGGGTACCACCTCGTCATTTCCAGCAATAAGTTTAGGAGGAGTTTGGGAGACGACTTCGGGCAGTGAAGCGCCGCATAACCAGACAACCCGTTTATCTCTGCAAACGGAATTCGCACCGAGTTAGCGGCCCTGTTCGAGCAACTTCTACAGGCTTTGGCCACACGCTGTGAGAAATGGAGCCCACGTTTAGCAGATTTGTAGGTGGCAATAAGGGTGTCCATGCACAGCCTGCAACCGGTAGTTTGGTGTAACTGTAAATCGCTCAGCCGTAAGCATCGGGAAGTCCCCCCTACCTAACAGCAGGCCGTCGGTCGCGTTCGCAGGCACTATCGGCCACAATTGGTCGTTAGAAAGCAGGCATTTATAGTTTCAATTAAAGAGGCCTCTTTACCAGTCAATACTCTCACGGAGGTCAAACTACAGCGTACCATCATCAGAGACTGTGAGGCCCATCGTTAGCCGCGGCTTACACGCACATCCGCCACATAGTCTAATAAGGCTTCACTCAACATGAACCTCGAGAATACCTTGAAGACTGCTTATTCTCTATAGTAGCGCACATTTCCGAAGTGGACCCCATGACATACTTCGAAGGTGGGTGCACCTGTATTTCCTCGTAGATGTGCGCGTAGTCAGGTTGGCAGCAACTTACTCGGGCCCTTT
|
| 73 |
+
>seq_36
|
| 74 |
+
CAGGCGATTGCTGCAGGCCAAAGCTGTTTCTAGGCGGCTACCGCGGGTGTAACCAGGCGCAGGGCAATACTATTTTTGGCAGCTCAGACTCGTATGCTTGTAGCAGCAGGAAATGGTGGAACTTCTGGTCCATTAAGACACTTGTCAGGAGTCGTCTATCTGGCTGTGAAAGGATGTATCAAACTCTATGTCTAAATACATCTTCTGGTGGGGGGCTATAAACCCATGCATTGGGTACACGAAAAATCTTTTTCACCCTACGATATCAGACTTTACGTTCGGTTTTGAGCCGCCTCGAGGGATTGAAGAAAAAAATCATCTGATGGCGGATCGAATAAGGCGCACACGAGCAATCGTATCCCTAAGTGCACCGTTCCATTAAGAAGAAGCTAATTGGTTACCGACGGTTACGTGTGATCATGCTCGTGCGGCATGGGTATAAGAGGGGAACGCGGCCCCTGTCCCACGTGACAGGTATAGTGGTTTAGAGGAGTGACTTCTCAAAAAATAAAATATCGAGATTGGCATGTCTTAAAACGGTCAGTGGGGTGACTGCGTTTGGAGATAAGTGGAATGGTCCGGCCTAATATGTAGCATAGCAATAGCTACCGGGGACCTGTTCCTTCGATTAACCAGATCCTATTCGTATCCAGTTCAGAGGCATAGGCAATCAATGATTTGCTGACAGCGTCAGGAGTCAAACAGAGCCGCCATGCGAACCCTGAGACGCCGCGGATAAGACAAACAATGGGGCCCATGTATTAGCGGACCCCTGGTAGACTGGCAACTATTCAAGTAGGATGTCACTTTAAATAAAGGTGCAGCCAGGAGTCGATTCTGGTAATTTGGACACTTAGACCCTATTGTGGAGCCGAGGTGTGGTCCCCTTAGCACCATTAGACGATGCGTATTAAGGGCATAATGTGTCAGTCACGGATGCAGATGCCTATAATAATTGCGATGCGGCGATACCACGTTCTCAGAATCTTACCTCGAGTGA
|
| 75 |
+
>seq_37
|
| 76 |
+
ATTCCGGCGGAGTTTACACCAGGTGTATCACAATTGGCGCGAGCCGGTTCATTTCAAGACCGTTCGCTTCTGACGCCGCTCCTGGATTCGGTGGCGCAAGGGACTGGGGTTCTGATTATACACGGTAGTCCCATCGGCTTCATAATGTCTCTATGGGGGATGTTCCGGACGCGGTCCACCTCTGACTAAGGCTCCCTCTTGTTACGTGAGAGCTAGCAGGCTTCCGGCTCGTTCGCGTCACGTATACATTAGCGAGAAGTTGGGGTTGGAACGTTTTCGACGAGTATCTGGAAGCCACATTCATAAAATGGACACGAGTACAAGACTCGGAGCGTGGGGATTCATATGCCTGGGGTCCGGCTCTAGCCGCGCGTTCCCGTCATCTGAAGTTTCACACTGTATTGCTAGACACTTCGTAATATAAACCGCTTGGTATACTGTTGTCCGGCACGTAGGGCATGAAAGCCCTTAAGCGTTCTGGCTGACGTGCTCGACAGCCGCCCGTCCTTCACATTTTCTTGGTGAGGCGGCACCTTTACTCTTTGCGAATATACAATCACACGCGGCCTGGAGGCCGGGAGCAAACAACGCAATACATGGGAGTCCCAATTACGTAGTCACCGGCATGCGGATGGTGACTGGTAAACGTTCGGTGTTCCCATCGGTCACACTTCTTTCACCCGGACCATACTACTTAACATTAGACGCTCAGGTGCAAAAAACCTGAGCTCAATCCTGGACTGGGTTGGACGAATATCCCTTGTACCACATGAGTCTTGGCGTCGCAAGCCAATGTTCCACCAAGCGTAAGCCTATCTCATGAGATCATAACCCCATTGCTCCAAAGAGACGTGTTGCACGGCAAAACTCTGATAGCTATCCACCCTGACGGGATGGGGATGCTTAGAGACGACTGAGTCGCTCTAGTCAACGGGTTCCGATCTTTATCGTTTGTCAGTCTCCCGTCTCAGAGGAAGCCGGTAGGACTAGGTGACCGGCA
|
| 77 |
+
>seq_38
|
| 78 |
+
GGCAAGTTTGCCCAATTACCGTATGCCTTACTGCTGTCAATAGGGCTTCGAATCGAACGACATTGCCGGACGTCTCTAGACCCCATTTTTTATCAGCCCAGGCGAAGGGTGTACAACGGTTGACCCAAAGTTGGAGGTAAAAAGAGGGCGGTTAGACTTTCAATCAACGAAATGCTTCAGGGAACCTATATTATCCCTATTTTCCAGAATCGGAGTCGGCCCCCTACCGCTTCGCTGGGACCTGGTCGTGTCATCTCCTCTCCAGCTTACCAGCGCAGGCAGCCAGCACACCGGAACAGTTAAACGGATACTGAGCTTGACATGCGTGGTTTGTGTCGTCCAAAGCCGGATAGTATAAGTTTAGCCGCATATAGGCGACCTCAGTGCTGATTGCCGATAAGGTTCTAGAGTCAATTTCAAACTCAAAAAACCCCGGAATAGGACATAACGTGTAGGGCACCCGGTCTAGAAGCTGTGCCACGCATACAGTTGGGTCTTACATCGTCTGAATACCCGATCACCTGGTCGACAAAACTTGGTGAATCGGTATGTGAGATAAGCCGCGGAAATGTGTGTCGTAATCCCAGACCTAGCAGTGGAGAGTAGCCCCACTAGGGCTACTATTAAGGTCGTTGCTTAACCCGCGAGCGATGCGGGATTGGTGGTGTTTAGACGACCGCGCTCCGTATTTTGGTGCAATCAATACCACACGGCAAGGGATGGTCGCGATACGCCAAGTGGTTGCCCGCCCTGTTCGCAGCCTCCTAGTTATACTCTGGGGGTAACCTAGCTAATGATGTGAATATTCACGCTTAACCCGTGATAAATTAAAAGGCCCCGCATGGCACAGAACCTATTAGGCTGGCTCAGTTCGTGCAAAATGAGCACATTCAGTCACACCGCAATAAATCATGGAATTTGGCACGTCCCTATGGGGCGTACTTAATATATAGGCTGTAAAACGCTTACGGGGCATATGTCCTTGCTGTTAATTGGCGAA
|
| 79 |
+
>seq_39
|
| 80 |
+
AATTGGTACCGTTGTGGGGGGCGTTGCAGCTGACGCAACATATATAAAGGCAGGTTCTTAAATTGCTCTACCGCCGAAACCGCATGGCTCGCCACTCAGCCGGTGCTATTACCCGCTAGCAATCTCTGCTTAGCGGGACGACCTAAATGGTGTTCGAGTTACGGAAAACCCGCAAAGAGGTACCATGCCGGACAGTATTTATCCAAATGTCCTATTAGCCACCACCCGAGAGGTTGATAGCATGTCTGCGAAGCTCCTCGTTTTCCGAGAGCTCCTGACACTCTTTTGCATCGGTTTAGATCAGAAAGACCTAGTACAATATACCCGGCTGCGTCCCATGAAGGATGGCGCTCGGCTACTGAGAGCCCTGGCGGAGGATGTAAAATCATTGGAATACCAGCAAACATGGGTTAATGCACCGGTCGGTTTGGCAGGGTTCCTGCGTCCCATGAATTCGGCCGTGACACGTCGCGTGATGGCCACAACTCACTCTTTTCATTACACTTTGATTGGGGCTGGATATACTTGGCGGTGAAGGACGCCGGATAAAGTGACTATAGACAATCAACGAGTACAGCGATGCAACCTCCACACCCAACCAAACGAGTTAGTGAACTCGTAGATATGCCAGCGCAACGTCATGTGGATATTCAATATGTCTTACTACAACACTTTTGCGCATTATCTCGGCCGTTCGCGTGTTAATATTGTAGTATCCAGACTACCGCTGTGGTTCTCACGAATACTTCAGAGGTGTCGACTCAGCATACGTGTATTAGGCCCTGGCCGTCCCTGGGTTTACGGACGGTCACTAACTAATACGCTGATGAGAGTAACGTCCCCCGCACGTAACCGTCATTAAGACACAGGAGAGTTTGTCATCCTACTAGGTTCACCGCAACGATCGCTCTCCCCTCGTGGACGTTTCCAACTCGAGGAGGGGAGATTACCTGAAAGTTCATTATAACCGGCTCTTGAGAGACGATTTGATGGGCGGTTG
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/data/task_info.json
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n_sequences": 40,
|
| 3 |
+
"sequence_length": 1000,
|
| 4 |
+
"motif_length": 30,
|
| 5 |
+
"mismatches": 13,
|
| 6 |
+
"alphabet": "ACGT",
|
| 7 |
+
"model": {
|
| 8 |
+
"name": "planted (l, d)-motif",
|
| 9 |
+
"background": "every base outside the sites is drawn independently and uniformly from A, C, G, T",
|
| 10 |
+
"sites": "each sequence contains exactly one binding site; it is the only length-l window of that sequence within Hamming distance d of the motif, and its distance to the motif is exactly d",
|
| 11 |
+
"positions": "site starts are uniform over 0 .. sequence_length - motif_length"
|
| 12 |
+
},
|
| 13 |
+
"output_schema": {
|
| 14 |
+
"sites_table": "two columns seq_id,start - one row per sequence in file order",
|
| 15 |
+
"motif_file": "one line with the motif (motif_length characters from ACGT)",
|
| 16 |
+
"analysis.py": "deterministic script regenerating both outputs from data/ within 20 minutes on one CPU"
|
| 17 |
+
}
|
| 18 |
+
}
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/instance_meta.json
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"params_used": {
|
| 3 |
+
"seed": 31415
|
| 4 |
+
},
|
| 5 |
+
"input_files": [
|
| 6 |
+
"data/promoters.fa",
|
| 7 |
+
"data/task_info.json"
|
| 8 |
+
],
|
| 9 |
+
"reference_files": [
|
| 10 |
+
"sites_ref.csv",
|
| 11 |
+
"motif_ref.txt",
|
| 12 |
+
"motif_ref.json"
|
| 13 |
+
],
|
| 14 |
+
"generation_time_seconds": 0.064
|
| 15 |
+
}
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/prompt_b1.md
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Discover a subtle transcription-factor binding motif
|
| 2 |
+
|
| 3 |
+
A transcription factor binds a degenerate DNA motif. You are given 40 promoter-like DNA sequences of 1000 bp, each of which contains exactly one binding site of this factor. Recover the consensus motif and the position of the binding site in every sequence.
|
| 4 |
+
|
| 5 |
+
The model (planted (l, d)-motif model): every base outside the binding sites is drawn independently and uniformly from `A, C, G, T`. Each of the 40 sequences contains exactly one binding site of length 30; the site differs from the consensus motif `M` (a string of length 30) at exactly 13 positions, and it is the only window of length 30 in its sequence within Hamming distance 13 of `M`. Site starts are uniform over `0 … 970`. Under this model the consensus motif is, with overwhelming probability, the only length-30 string that has a window within distance 13 in every sequence, so the motif and every site position are uniquely determined by the data, and a candidate answer can be checked directly: every sequence must contain exactly one window within distance 13 of the reported motif, at the reported start.
|
| 6 |
+
|
| 7 |
+
## Background and method (full procedure)
|
| 8 |
+
|
| 9 |
+
Sites are very degenerate: 13 of 30 positions differ from the consensus, so two sites typically agree at only about 38% of positions, barely more than two random windows (25%). This is the "subtle motif" regime (Keich & Pevzner 2002) in which profile methods (EM/MEME, Gibbs sampling), random projection and local search started from scratch fail, even though the motif is uniquely identifiable. What makes the problem tractable here is prior knowledge: the following 16 binding sites were **confirmed experimentally** (for example by ChIP-exo footprints):
|
| 10 |
+
|
| 11 |
+
| sequence | site start (0-based) |
|
| 12 |
+
|---|---|
|
| 13 |
+
| seq_00 | 864 |
|
| 14 |
+
| seq_01 | 60 |
|
| 15 |
+
| seq_02 | 513 |
|
| 16 |
+
| seq_10 | 745 |
|
| 17 |
+
| seq_11 | 323 |
|
| 18 |
+
| seq_12 | 632 |
|
| 19 |
+
| seq_13 | 599 |
|
| 20 |
+
| seq_14 | 348 |
|
| 21 |
+
| seq_15 | 591 |
|
| 22 |
+
| seq_16 | 353 |
|
| 23 |
+
| seq_17 | 122 |
|
| 24 |
+
| seq_18 | 44 |
|
| 25 |
+
| seq_26 | 315 |
|
| 26 |
+
| seq_29 | 340 |
|
| 27 |
+
| seq_30 | 456 |
|
| 28 |
+
| seq_35 | 829 |
|
| 29 |
+
|
| 30 |
+
1. Parse the FASTA file into an integer array `(n_sequences, sequence_length)`; build all windows of length 30 (a sliding-window view of shape `(n, L - l + 1, l)`).
|
| 31 |
+
2. **Seed consensus.** Stack the 16 confirmed sites and take the column-wise majority base; each column of this seed is correct with high probability because each site keeps the consensus base at ~57% of positions.
|
| 32 |
+
3. **Consensus refinement.** Repeat until the motif stops changing: for every sequence take the window with the smallest Hamming distance to the current motif, and recompute the motif as the column-wise majority of those windows.
|
| 33 |
+
4. **Local search on the combinatorial objective.** Maximise, lexicographically, (number of sequences that contain a window within distance 13 of the motif, minus the sum over sequences of the minimum window distance) by steepest-ascent single-base substitutions (for each of the 30 positions and 3 alternative bases, update the distance matrix incrementally). If the optimum still leaves some sequence without a window within distance 13, apply a random kick (change 3 random positions of the best motif) and repeat the local search; keep the best motif and stop as soon as the certificate holds.
|
| 34 |
+
5. **Certificate and output.** The motif is correct when every sequence has exactly one window within distance 13 and every confirmed site is at distance exactly 13; report that window's start for each sequence and the motif. Use a fixed random seed so `analysis.py` is deterministic.
|
| 35 |
+
|
| 36 |
+
## Input
|
| 37 |
+
|
| 38 |
+
- `data/promoters.fa` — FASTA text file with exactly 40 records. Each record is a header line `>seq_NN` (`seq_00` … `seq_39`, in this order) followed by exactly one sequence line of 1000 characters from `A`, `C`, `G`, `T` (no line wrapping, no other characters). Positions are 0-based.
|
| 39 |
+
- `data/task_info.json` — UTF-8 JSON object; all fields are required and always present (none optional):
|
| 40 |
+
- `n_sequences` (integer, = 40), `sequence_length` (integer, = 1000), `motif_length` (integer, = 30), `mismatches` (integer, = 13), `alphabet` (string, `"ACGT"`);
|
| 41 |
+
- `model` (object with string fields `name`, `background`, `sites`, `positions`) — a restatement of the model described above;
|
| 42 |
+
- `output_schema` (object with string fields `sites_table`, `motif_file`, `analysis.py`) — a human-readable restatement of the output contract below; the scorer, not this text, is authoritative.
|
| 43 |
+
|
| 44 |
+
No other file is provided; there is no annotation, no expression data and no hidden parameter.
|
| 45 |
+
|
| 46 |
+
## Output
|
| 47 |
+
|
| 48 |
+
Write, relative to the working directory:
|
| 49 |
+
|
| 50 |
+
1. `results/sites.csv`
|
| 51 |
+
Exact columns:
|
| 52 |
+
`seq_id,start`
|
| 53 |
+
One row per sequence, in file order `seq_00` … `seq_39` (exactly 40 rows). `seq_id` is the FASTA identifier (string); `start` is the 0-based start of the binding site in that sequence, a plain non-negative integer in `[0, 970]`. Header row exactly `seq_id,start`; no index column, no quotes, no extra columns; UTF-8, comma-separated.
|
| 54 |
+
|
| 55 |
+
2. `results/motif.txt`
|
| 56 |
+
One line containing the consensus motif: exactly 30 characters from `ACGT`.
|
| 57 |
+
|
| 58 |
+
3. `analysis.py`
|
| 59 |
+
A single self-contained Python 3.11+ script (standard library, `numpy`, `scipy` only; no network; no `importlib` / `__import__`; no reference to any `reference/` path) that, run as `python analysis.py` from the working directory, reads only `data/` and regenerates both result files (creating `results/` if needed). It must be deterministic (fix every random seed) and finish within 20 minutes on one CPU core. The evaluator re-executes it in a clean copy of the workspace and scores the regenerated files as well as the submitted ones; the lower of the two scores counts.
|
| 60 |
+
|
| 61 |
+
Scoring: `70 × (fraction of sequences whose reported start equals the true site start) + 30 × [reported motif equals the true motif]`. Both are uniquely determined by the data under the stated model. A contract violation zeroes the affected file's component; a modified data file or an `analysis.py` that fails to reproduce the outputs scores 0.
|
| 62 |
+
|
| 63 |
+
Time budget: plan for at most about 60 minutes of wall-clock work in total, including any searches your script runs. Do not launch open-ended or exhaustive searches. If you cannot certify the motif within that budget, write your best answer to both result files, make sure `analysis.py` regenerates it deterministically, and stop — partially correct site positions receive partial credit, whereas an unfinished workspace scores 0.
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/prompt_b2.md
ADDED
|
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Discover a subtle transcription-factor binding motif
|
| 2 |
+
|
| 3 |
+
A transcription factor binds a degenerate DNA motif. You are given 40 promoter-like DNA sequences of 1000 bp, each of which contains exactly one binding site of this factor. Recover the consensus motif and the position of the binding site in every sequence.
|
| 4 |
+
|
| 5 |
+
The model (planted (l, d)-motif model): every base outside the binding sites is drawn independently and uniformly from `A, C, G, T`. Each of the 40 sequences contains exactly one binding site of length 30; the site differs from the consensus motif `M` (a string of length 30) at exactly 13 positions, and it is the only window of length 30 in its sequence within Hamming distance 13 of `M`. Site starts are uniform over `0 … 970`. Under this model the consensus motif is, with overwhelming probability, the only length-30 string that has a window within distance 13 in every sequence, so the motif and every site position are uniquely determined by the data, and a candidate answer can be checked directly: every sequence must contain exactly one window within distance 13 of the reported motif, at the reported start.
|
| 6 |
+
|
| 7 |
+
The sites are very degenerate (the subtle-motif regime), but 8 binding sites were **confirmed experimentally**:
|
| 8 |
+
|
| 9 |
+
| sequence | site start (0-based) |
|
| 10 |
+
|---|---|
|
| 11 |
+
| seq_00 | 864 |
|
| 12 |
+
| seq_01 | 60 |
|
| 13 |
+
| seq_02 | 513 |
|
| 14 |
+
| seq_11 | 323 |
|
| 15 |
+
| seq_12 | 632 |
|
| 16 |
+
| seq_14 | 348 |
|
| 17 |
+
| seq_15 | 591 |
|
| 18 |
+
| seq_30 | 456 |
|
| 19 |
+
|
| 20 |
+
Use them: seed a consensus from the confirmed sites, refine it by re-aligning each sequence to its closest window, and finish with a combinatorial local search (with random restarts/kicks) that maximises the number of sequences containing a window within distance 13 of the motif; stop when the certificate holds for every sequence.
|
| 21 |
+
|
| 22 |
+
## Input
|
| 23 |
+
|
| 24 |
+
- `data/promoters.fa` — FASTA text file with exactly 40 records. Each record is a header line `>seq_NN` (`seq_00` … `seq_39`, in this order) followed by exactly one sequence line of 1000 characters from `A`, `C`, `G`, `T` (no line wrapping, no other characters). Positions are 0-based.
|
| 25 |
+
- `data/task_info.json` — UTF-8 JSON object; all fields are required and always present (none optional):
|
| 26 |
+
- `n_sequences` (integer, = 40), `sequence_length` (integer, = 1000), `motif_length` (integer, = 30), `mismatches` (integer, = 13), `alphabet` (string, `"ACGT"`);
|
| 27 |
+
- `model` (object with string fields `name`, `background`, `sites`, `positions`) — a restatement of the model described above;
|
| 28 |
+
- `output_schema` (object with string fields `sites_table`, `motif_file`, `analysis.py`) — a human-readable restatement of the output contract below; the scorer, not this text, is authoritative.
|
| 29 |
+
|
| 30 |
+
No other file is provided; there is no annotation, no expression data and no hidden parameter.
|
| 31 |
+
|
| 32 |
+
## Output
|
| 33 |
+
|
| 34 |
+
Write, relative to the working directory:
|
| 35 |
+
|
| 36 |
+
1. `results/sites.csv`
|
| 37 |
+
Exact columns:
|
| 38 |
+
`seq_id,start`
|
| 39 |
+
One row per sequence, in file order `seq_00` … `seq_39` (exactly 40 rows). `seq_id` is the FASTA identifier (string); `start` is the 0-based start of the binding site in that sequence, a plain non-negative integer in `[0, 970]`. Header row exactly `seq_id,start`; no index column, no quotes, no extra columns; UTF-8, comma-separated.
|
| 40 |
+
|
| 41 |
+
2. `results/motif.txt`
|
| 42 |
+
One line containing the consensus motif: exactly 30 characters from `ACGT`.
|
| 43 |
+
|
| 44 |
+
3. `analysis.py`
|
| 45 |
+
A single self-contained Python 3.11+ script (standard library, `numpy`, `scipy` only; no network; no `importlib` / `__import__`; no reference to any `reference/` path) that, run as `python analysis.py` from the working directory, reads only `data/` and regenerates both result files (creating `results/` if needed). It must be deterministic (fix every random seed) and finish within 20 minutes on one CPU core. The evaluator re-executes it in a clean copy of the workspace and scores the regenerated files as well as the submitted ones; the lower of the two scores counts.
|
| 46 |
+
|
| 47 |
+
Scoring: `70 × (fraction of sequences whose reported start equals the true site start) + 30 × [reported motif equals the true motif]`. Both are uniquely determined by the data under the stated model. A contract violation zeroes the affected file's component; a modified data file or an `analysis.py` that fails to reproduce the outputs scores 0.
|
| 48 |
+
|
| 49 |
+
Time budget: plan for at most about 60 minutes of wall-clock work in total, including any searches your script runs. Do not launch open-ended or exhaustive searches. If you cannot certify the motif within that budget, write your best answer to both result files, make sure `analysis.py` regenerates it deterministically, and stop — partially correct site positions receive partial credit, whereas an unfinished workspace scores 0.
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/prompt_b3.md
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Discover a subtle transcription-factor binding motif
|
| 2 |
+
|
| 3 |
+
A transcription factor binds a degenerate DNA motif. You are given 40 promoter-like DNA sequences of 1000 bp, each of which contains exactly one binding site of this factor. Recover the consensus motif and the position of the binding site in every sequence.
|
| 4 |
+
|
| 5 |
+
The model (planted (l, d)-motif model): every base outside the binding sites is drawn independently and uniformly from `A, C, G, T`. Each of the 40 sequences contains exactly one binding site of length 30; the site differs from the consensus motif `M` (a string of length 30) at exactly 13 positions, and it is the only window of length 30 in its sequence within Hamming distance 13 of `M`. Site starts are uniform over `0 … 970`. Under this model the consensus motif is, with overwhelming probability, the only length-30 string that has a window within distance 13 in every sequence, so the motif and every site position are uniquely determined by the data, and a candidate answer can be checked directly: every sequence must contain exactly one window within distance 13 of the reported motif, at the reported start.
|
| 6 |
+
|
| 7 |
+
## Input
|
| 8 |
+
|
| 9 |
+
- `data/promoters.fa` — FASTA text file with exactly 40 records. Each record is a header line `>seq_NN` (`seq_00` … `seq_39`, in this order) followed by exactly one sequence line of 1000 characters from `A`, `C`, `G`, `T` (no line wrapping, no other characters). Positions are 0-based.
|
| 10 |
+
- `data/task_info.json` — UTF-8 JSON object; all fields are required and always present (none optional):
|
| 11 |
+
- `n_sequences` (integer, = 40), `sequence_length` (integer, = 1000), `motif_length` (integer, = 30), `mismatches` (integer, = 13), `alphabet` (string, `"ACGT"`);
|
| 12 |
+
- `model` (object with string fields `name`, `background`, `sites`, `positions`) — a restatement of the model described above;
|
| 13 |
+
- `output_schema` (object with string fields `sites_table`, `motif_file`, `analysis.py`) — a human-readable restatement of the output contract below; the scorer, not this text, is authoritative.
|
| 14 |
+
|
| 15 |
+
No other file is provided; there is no annotation, no expression data and no hidden parameter.
|
| 16 |
+
|
| 17 |
+
## Output
|
| 18 |
+
|
| 19 |
+
Write, relative to the working directory:
|
| 20 |
+
|
| 21 |
+
1. `results/sites.csv`
|
| 22 |
+
Exact columns:
|
| 23 |
+
`seq_id,start`
|
| 24 |
+
One row per sequence, in file order `seq_00` … `seq_39` (exactly 40 rows). `seq_id` is the FASTA identifier (string); `start` is the 0-based start of the binding site in that sequence, a plain non-negative integer in `[0, 970]`. Header row exactly `seq_id,start`; no index column, no quotes, no extra columns; UTF-8, comma-separated.
|
| 25 |
+
|
| 26 |
+
2. `results/motif.txt`
|
| 27 |
+
One line containing the consensus motif: exactly 30 characters from `ACGT`.
|
| 28 |
+
|
| 29 |
+
3. `analysis.py`
|
| 30 |
+
A single self-contained Python 3.11+ script (standard library, `numpy`, `scipy` only; no network; no `importlib` / `__import__`; no reference to any `reference/` path) that, run as `python analysis.py` from the working directory, reads only `data/` and regenerates both result files (creating `results/` if needed). It must be deterministic (fix every random seed) and finish within 20 minutes on one CPU core. The evaluator re-executes it in a clean copy of the workspace and scores the regenerated files as well as the submitted ones; the lower of the two scores counts.
|
| 31 |
+
|
| 32 |
+
Scoring: `70 × (fraction of sequences whose reported start equals the true site start) + 30 × [reported motif equals the true motif]`. Both are uniquely determined by the data under the stated model. A contract violation zeroes the affected file's component; a modified data file or an `analysis.py` that fails to reproduce the outputs scores 0.
|
| 33 |
+
|
| 34 |
+
Time budget: plan for at most about 60 minutes of wall-clock work in total, including any searches your script runs. Do not launch open-ended or exhaustive searches. If you cannot certify the motif within that budget, write your best answer to both result files, make sure `analysis.py` regenerates it deterministically, and stop — partially correct site positions receive partial credit, whereas an unfinished workspace scores 0.
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/prompt_b4.md
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Discover a subtle transcription-factor binding motif
|
| 2 |
+
|
| 3 |
+
A transcription factor binds a degenerate DNA motif. You are given 40 promoter-like DNA sequences of 1000 bp, each of which contains exactly one binding site of this factor. Recover the consensus motif and the position of the binding site in every sequence.
|
| 4 |
+
|
| 5 |
+
The model (planted (l, d)-motif model): every base outside the binding sites is drawn independently and uniformly from `A, C, G, T`. Each of the 40 sequences contains exactly one binding site of length 30; the site differs from the consensus motif `M` (a string of length 30) at exactly 13 positions, and it is the only window of length 30 in its sequence within Hamming distance 13 of `M`. Site starts are uniform over `0 … 970`. Under this model the consensus motif is, with overwhelming probability, the only length-30 string that has a window within distance 13 in every sequence, so the motif and every site position are uniquely determined by the data, and a candidate answer can be checked directly: every sequence must contain exactly one window within distance 13 of the reported motif, at the reported start.
|
| 6 |
+
|
| 7 |
+
Some background from the transcription-factor biochemistry team, forwarded verbatim because I was asked to pass along everything. The sequences were exported on a Friday from an internal browser session; the export tool assigns the identifiers seq_00, seq_01, ... in the order the regions were selected, which carries no information about where the sites are. The overall GC content of the file is 50.0% and every sequence is exactly 1000 bp with no ambiguous bases (no N). The regions were chosen by a colleague for a different project and re-used here; their genomic coordinates, gene annotations and expression data are not included because they are not relevant to this analysis. Someone suggested plotting a sequence logo of the final answer, which is a nice habit but not a requirement. The workstation is shared, so please keep the runtime of the final script reasonable, although nothing in the deliverables depends on runtime as long as the limit in the output contract is respected.
|
| 8 |
+
|
| 9 |
+
## Input
|
| 10 |
+
|
| 11 |
+
- `data/promoters.fa` — FASTA text file with exactly 40 records. Each record is a header line `>seq_NN` (`seq_00` … `seq_39`, in this order) followed by exactly one sequence line of 1000 characters from `A`, `C`, `G`, `T` (no line wrapping, no other characters). Positions are 0-based.
|
| 12 |
+
- `data/task_info.json` — UTF-8 JSON object; all fields are required and always present (none optional):
|
| 13 |
+
- `n_sequences` (integer, = 40), `sequence_length` (integer, = 1000), `motif_length` (integer, = 30), `mismatches` (integer, = 13), `alphabet` (string, `"ACGT"`);
|
| 14 |
+
- `model` (object with string fields `name`, `background`, `sites`, `positions`) — a restatement of the model described above;
|
| 15 |
+
- `output_schema` (object with string fields `sites_table`, `motif_file`, `analysis.py`) — a human-readable restatement of the output contract below; the scorer, not this text, is authoritative.
|
| 16 |
+
|
| 17 |
+
No other file is provided; there is no annotation, no expression data and no hidden parameter.
|
| 18 |
+
|
| 19 |
+
## Output
|
| 20 |
+
|
| 21 |
+
Write, relative to the working directory:
|
| 22 |
+
|
| 23 |
+
1. `results/sites.csv`
|
| 24 |
+
Exact columns:
|
| 25 |
+
`seq_id,start`
|
| 26 |
+
One row per sequence, in file order `seq_00` … `seq_39` (exactly 40 rows). `seq_id` is the FASTA identifier (string); `start` is the 0-based start of the binding site in that sequence, a plain non-negative integer in `[0, 970]`. Header row exactly `seq_id,start`; no index column, no quotes, no extra columns; UTF-8, comma-separated.
|
| 27 |
+
|
| 28 |
+
2. `results/motif.txt`
|
| 29 |
+
One line containing the consensus motif: exactly 30 characters from `ACGT`.
|
| 30 |
+
|
| 31 |
+
3. `analysis.py`
|
| 32 |
+
A single self-contained Python 3.11+ script (standard library, `numpy`, `scipy` only; no network; no `importlib` / `__import__`; no reference to any `reference/` path) that, run as `python analysis.py` from the working directory, reads only `data/` and regenerates both result files (creating `results/` if needed). It must be deterministic (fix every random seed) and finish within 20 minutes on one CPU core. The evaluator re-executes it in a clean copy of the workspace and scores the regenerated files as well as the submitted ones; the lower of the two scores counts.
|
| 33 |
+
|
| 34 |
+
Scoring: `70 × (fraction of sequences whose reported start equals the true site start) + 30 × [reported motif equals the true motif]`. Both are uniquely determined by the data under the stated model. A contract violation zeroes the affected file's component; a modified data file or an `analysis.py` that fails to reproduce the outputs scores 0.
|
| 35 |
+
|
| 36 |
+
Time budget: plan for at most about 60 minutes of wall-clock work in total, including any searches your script runs. Do not launch open-ended or exhaustive searches. If you cannot certify the motif within that budget, write your best answer to both result files, make sure `analysis.py` regenerates it deterministically, and stop — partially correct site positions receive partial credit, whereas an unfinished workspace scores 0.
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/reference/motif_ref.json
ADDED
|
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"motif": "ACTCTGTCATTATTCACCATGGGTCTCTAT",
|
| 3 |
+
"motif_length": 30,
|
| 4 |
+
"mismatches": 13,
|
| 5 |
+
"n_sequences": 40,
|
| 6 |
+
"sequence_length": 1000,
|
| 7 |
+
"sites": [
|
| 8 |
+
864,
|
| 9 |
+
60,
|
| 10 |
+
513,
|
| 11 |
+
897,
|
| 12 |
+
109,
|
| 13 |
+
346,
|
| 14 |
+
85,
|
| 15 |
+
515,
|
| 16 |
+
445,
|
| 17 |
+
168,
|
| 18 |
+
745,
|
| 19 |
+
323,
|
| 20 |
+
632,
|
| 21 |
+
599,
|
| 22 |
+
348,
|
| 23 |
+
591,
|
| 24 |
+
353,
|
| 25 |
+
122,
|
| 26 |
+
44,
|
| 27 |
+
642,
|
| 28 |
+
926,
|
| 29 |
+
82,
|
| 30 |
+
667,
|
| 31 |
+
725,
|
| 32 |
+
799,
|
| 33 |
+
768,
|
| 34 |
+
315,
|
| 35 |
+
537,
|
| 36 |
+
220,
|
| 37 |
+
340,
|
| 38 |
+
456,
|
| 39 |
+
586,
|
| 40 |
+
734,
|
| 41 |
+
833,
|
| 42 |
+
671,
|
| 43 |
+
829,
|
| 44 |
+
834,
|
| 45 |
+
912,
|
| 46 |
+
378,
|
| 47 |
+
271
|
| 48 |
+
],
|
| 49 |
+
"data_sha256": {
|
| 50 |
+
"promoters.fa": "ca7047c3962c14663d2e9df05ef86d456fc057b783574fe47a5e14311b4f7b17",
|
| 51 |
+
"task_info.json": "8a9a0bb1b2871079226846a50933eacf453010432fdd5668028fe838d0e88417"
|
| 52 |
+
},
|
| 53 |
+
"confirmed_b1": [
|
| 54 |
+
0,
|
| 55 |
+
1,
|
| 56 |
+
2,
|
| 57 |
+
10,
|
| 58 |
+
11,
|
| 59 |
+
12,
|
| 60 |
+
13,
|
| 61 |
+
14,
|
| 62 |
+
15,
|
| 63 |
+
16,
|
| 64 |
+
17,
|
| 65 |
+
18,
|
| 66 |
+
26,
|
| 67 |
+
29,
|
| 68 |
+
30,
|
| 69 |
+
35
|
| 70 |
+
],
|
| 71 |
+
"confirmed_b2": [
|
| 72 |
+
0,
|
| 73 |
+
1,
|
| 74 |
+
2,
|
| 75 |
+
11,
|
| 76 |
+
12,
|
| 77 |
+
14,
|
| 78 |
+
15,
|
| 79 |
+
30
|
| 80 |
+
]
|
| 81 |
+
}
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/reference/motif_ref.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
ACTCTGTCATTATTCACCATGGGTCTCTAT
|
tasks/biology.subtle_regulatory_motif_discovery__seed31415/reference/sites_ref.csv
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
seq_id,start
|
| 2 |
+
seq_00,864
|
| 3 |
+
seq_01,60
|
| 4 |
+
seq_02,513
|
| 5 |
+
seq_03,897
|
| 6 |
+
seq_04,109
|
| 7 |
+
seq_05,346
|
| 8 |
+
seq_06,85
|
| 9 |
+
seq_07,515
|
| 10 |
+
seq_08,445
|
| 11 |
+
seq_09,168
|
| 12 |
+
seq_10,745
|
| 13 |
+
seq_11,323
|
| 14 |
+
seq_12,632
|
| 15 |
+
seq_13,599
|
| 16 |
+
seq_14,348
|
| 17 |
+
seq_15,591
|
| 18 |
+
seq_16,353
|
| 19 |
+
seq_17,122
|
| 20 |
+
seq_18,44
|
| 21 |
+
seq_19,642
|
| 22 |
+
seq_20,926
|
| 23 |
+
seq_21,82
|
| 24 |
+
seq_22,667
|
| 25 |
+
seq_23,725
|
| 26 |
+
seq_24,799
|
| 27 |
+
seq_25,768
|
| 28 |
+
seq_26,315
|
| 29 |
+
seq_27,537
|
| 30 |
+
seq_28,220
|
| 31 |
+
seq_29,340
|
| 32 |
+
seq_30,456
|
| 33 |
+
seq_31,586
|
| 34 |
+
seq_32,734
|
| 35 |
+
seq_33,833
|
| 36 |
+
seq_34,671
|
| 37 |
+
seq_35,829
|
| 38 |
+
seq_36,834
|
| 39 |
+
seq_37,912
|
| 40 |
+
seq_38,378
|
| 41 |
+
seq_39,271
|
tasks/biology.unseen_perturbation_rna_response__seed31415/data/README.md
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
Predict downstream measured RNA responses to held-out single-gene CRISPRi interventions in K562. All NPZ files are numeric/Unicode and must be loaded without pickle. See metadata.json for normalization, ordering, score and resource contracts. train_condition_stats.npz contains training-only means, per-cell sample variances and cell counts; validation_targets.npz contains measured validation GT-half means and scoring weights for tuning; test_queries.json contains identities only. All target Ensembl IDs occur as readout rows in genes.json. Their observed expression across training conditions is public information; their held-out intervention responses are not supplied. The directly targeted gene has zero scoring weight for its own condition. Training and test conditions are disjoint by canonical gene, and technical cell halves are not biological replicates.
|
tasks/biology.unseen_perturbation_rna_response__seed31415/data/genes.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tasks/biology.unseen_perturbation_rna_response__seed31415/data/manifest.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema_version": 1,
|
| 3 |
+
"files": {
|
| 4 |
+
"README.md": {
|
| 5 |
+
"sha256": "517dc558584be77eb2549d791a9784028c061a6e2998abea72897eee1c7e7b77",
|
| 6 |
+
"bytes": 868
|
| 7 |
+
},
|
| 8 |
+
"genes.json": {
|
| 9 |
+
"sha256": "b2b436ab503e8eb80d291bfc527dbad7777ecb27a7cf6806cad96fed1708559a",
|
| 10 |
+
"bytes": 153793
|
| 11 |
+
},
|
| 12 |
+
"metadata.json": {
|
| 13 |
+
"sha256": "6541388b6ad8270b791c909c3c09d1a548108838b9f0f969fef7ded7861cc04b",
|
| 14 |
+
"bytes": 3778
|
| 15 |
+
},
|
| 16 |
+
"provenance.json": {
|
| 17 |
+
"sha256": "8ede8971696cd9b9b58ae5767de4b64cc3c830e3012dd814f20b0647230f576e",
|
| 18 |
+
"bytes": 1905
|
| 19 |
+
},
|
| 20 |
+
"test_queries.json": {
|
| 21 |
+
"sha256": "c6142c4a549256e30994a28e1747b3ad1a76ac4c6a72ab7474138de2cb3d067a",
|
| 22 |
+
"bytes": 5423
|
| 23 |
+
},
|
| 24 |
+
"train_condition_stats.npz": {
|
| 25 |
+
"sha256": "213a9cfadf6122065a5f5f5beec43efd18c4ec70015e42e61d142a78fe5c2ae4",
|
| 26 |
+
"bytes": 54440658
|
| 27 |
+
},
|
| 28 |
+
"validation_targets.npz": {
|
| 29 |
+
"sha256": "d40483f6f1f22c35faa091d3150e163655827939fc53366565e948b72a872a72",
|
| 30 |
+
"bytes": 11697141
|
| 31 |
+
}
|
| 32 |
+
},
|
| 33 |
+
"public_test_content": "233canonicaltargetEnsemblIDs and requiredshape only; allreadoutgeneidentities",
|
| 34 |
+
"private_evaluation_arrays_included": false
|
| 35 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/data/metadata.json
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"dataset": "ReplogleWeissman2022_K562_essential",
|
| 3 |
+
"expression": "Mean of per-cell log1p(10000*RNAcounts/sum(all8563verifiedreleasedRNA))",
|
| 4 |
+
"target_semantics": "Unseen single canonical target ENSG; all guides/cells of each target assigned together",
|
| 5 |
+
"train_stats": "control row then1082trainingconditions; variance is per-cell sample variance (ddof1), n_cells counts cells; means have not used heldout perturbation responses",
|
| 6 |
+
"genes": "All output and target descriptors share exact genes.json gene_ids order; all1546eligibletargets are readout rows",
|
| 7 |
+
"validation": "Only validation GT-half means and GT-half-derived downstream metricweights are public for tuning; noTD",
|
| 8 |
+
"test": "Only233targetIDs and requested shape; no responses, technicalduplicate, weights, originalvarresponse summaries or uns",
|
| 9 |
+
"primary_metric": "Mean across conditions of sum_g(weight_g*(prediction_g-GT_g)^2); own targetedgene weight=0 and other weights sum1; weights derived from squared minmax abs Welch t versus split-controlGT cells",
|
| 10 |
+
"no_test_response_in_gene_descriptors": true,
|
| 11 |
+
"files": [
|
| 12 |
+
"README.md",
|
| 13 |
+
"genes.json",
|
| 14 |
+
"manifest.json",
|
| 15 |
+
"metadata.json",
|
| 16 |
+
"provenance.json",
|
| 17 |
+
"test_queries.json",
|
| 18 |
+
"train_condition_stats.npz",
|
| 19 |
+
"validation_targets.npz"
|
| 20 |
+
],
|
| 21 |
+
"output_shape": [
|
| 22 |
+
233,
|
| 23 |
+
3323
|
| 24 |
+
],
|
| 25 |
+
"task_id": "biology.unseen_perturbation_rna_response",
|
| 26 |
+
"schema_version": 1,
|
| 27 |
+
"cell_line": "K562",
|
| 28 |
+
"perturbation_type": "single-target CRISPR interference",
|
| 29 |
+
"normalization_denominator_features": 8563,
|
| 30 |
+
"feature_selection": "2000 highly variable RNA genes chosen using training target/control cells only, union all1546eligible targetRNA genes;3323total in original RNA-column order",
|
| 31 |
+
"scientific_scoring": {
|
| 32 |
+
"primary_loss": "mean over233conditions of sum over3323genes w*(prediction-measuredGTmean)^2",
|
| 33 |
+
"own_target_weight": "Exactly0 for the canonical target ENSG; remaining weights renormalizedsum1",
|
| 34 |
+
"weight_construction": "GT-half versus same-role controlGT-half Welch t with denominator sqrt(sample_variance/n+control_sample_variance/n); zero denominator t=0; minmax abs(t) over all3323genes, square, then self=0 and renormalize; zero remaining weights are an author data error",
|
| 35 |
+
"score_formula": "clip(100*(B-L)/(B-T),0,100)",
|
| 36 |
+
"B": 0.03441638519456471,
|
| 37 |
+
"T": 0.00714613515111071,
|
| 38 |
+
"B_definition": "Loss of equal-weight mean response of1082noncontroltrainingconditions",
|
| 39 |
+
"T_definition": "Loss of private technicalsecondhalfmeans against private GThalfmeans under same weights",
|
| 40 |
+
"GT_origin": "Private measured cell statistics, not an author model prediction",
|
| 41 |
+
"technical_half_limit": "Same experimental cohort, not independent biological replicates or a universal attainable predictor bound"
|
| 42 |
+
},
|
| 43 |
+
"output": {
|
| 44 |
+
"path": "results/predictions.npy",
|
| 45 |
+
"dtype": [
|
| 46 |
+
"float32",
|
| 47 |
+
"float64"
|
| 48 |
+
],
|
| 49 |
+
"shape": [
|
| 50 |
+
233,
|
| 51 |
+
3323
|
| 52 |
+
],
|
| 53 |
+
"finite": true,
|
| 54 |
+
"maximum_bytes": 8388608,
|
| 55 |
+
"source_path": "analysis.py",
|
| 56 |
+
"maximum_source_bytes": 65536,
|
| 57 |
+
"maximum_compressed_source_bytes": 12288,
|
| 58 |
+
"maximum_source_literal_characters": 8192,
|
| 59 |
+
"maximum_single_literal_characters": 1024,
|
| 60 |
+
"literal_definition": "Characters of all string, bytes, f-string, numeric and comment tokens in analysis.py; compressed size is zlib level 9 of the source bytes"
|
| 61 |
+
},
|
| 62 |
+
"resource_contract": {
|
| 63 |
+
"fresh_replay_seconds": 1200,
|
| 64 |
+
"CPUs": 4,
|
| 65 |
+
"threads": 4,
|
| 66 |
+
"memory_GiB": 8,
|
| 67 |
+
"development_budget": "Read timeout_seconds in task_info.json; configured independently of replay",
|
| 68 |
+
"scientific_data_access": "Only supplied public data; no external datasets, models or knowledge resources; provider connectivity is separate",
|
| 69 |
+
"replay_atol": 1e-05,
|
| 70 |
+
"replay_rtol": 1e-06
|
| 71 |
+
}
|
| 72 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/data/provenance.json
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema_version": 1,
|
| 3 |
+
"dataset": "Replogle et al. (2022) genome-scale Perturb-seq, K562 essential-gene CRISPRi, day 6",
|
| 4 |
+
"license": "CC-BY-4.0",
|
| 5 |
+
"original_paper": "https://doi.org/10.1016/j.cell.2022.05.013",
|
| 6 |
+
"feature_identity_audit": "All8563releasedfeatureEnsemblIDs uniquely match original Gene Expression annotation;0guide/nonRNA/engineeredcolumns retained. Original CellRanger table includes36601RNA+4561guidefeatures; releasedX was checked explicitly.",
|
| 7 |
+
"source_prefilter": "Released raw counts already restrict genes to mean>0.01UMI/cell. Our normalization sums all8563verifiedreleasedRNAgenes before panel selection.",
|
| 8 |
+
"target_identity": "obs.gene_id exact var.ensembl_id mapping;TBCE and HSPA14 resolved by ID, not symbol guessing.",
|
| 9 |
+
"eligible_cohort": "1546 target IDs with >=64 cells and unique measured RNA identity",
|
| 10 |
+
"split": "Fixed gene-disjoint assignment of canonical target IDs to training (1082), validation (231) and test (233); all guides/cells per target stay together. The assignment recipe is private.",
|
| 11 |
+
"technical_halves": "Each condition was divided once into two fixed technical cell halves; the first supplies GT statistics. The division recipe is private.",
|
| 12 |
+
"training_cells": 188047,
|
| 13 |
+
"training_controls": 7483,
|
| 14 |
+
"validation_controls": 1604,
|
| 15 |
+
"test_controls": 1604,
|
| 16 |
+
"public_numeric_source_hashes": {
|
| 17 |
+
"train_condition_stats.npz": "213a9cfadf6122065a5f5f5beec43efd18c4ec70015e42e61d142a78fe5c2ae4",
|
| 18 |
+
"validation_targets.npz": "d40483f6f1f22c35faa091d3150e163655827939fc53366565e948b72a872a72"
|
| 19 |
+
},
|
| 20 |
+
"offline_boundary": "Use only the supplied public files. Retrieving the original measurements or reconstructing the held-out split from external data breaches the declared holdout.",
|
| 21 |
+
"no_original_full_data_summary_priors": "Original var mean/std/cv/fano/ncounts/ncells and obs normalization summaries are not exported as predictive inputs."
|
| 22 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/data/test_queries.json
ADDED
|
@@ -0,0 +1,241 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"condition_ids": [
|
| 3 |
+
"ENSG00000188846",
|
| 4 |
+
"ENSG00000168028",
|
| 5 |
+
"ENSG00000075856",
|
| 6 |
+
"ENSG00000108654",
|
| 7 |
+
"ENSG00000149636",
|
| 8 |
+
"ENSG00000198700",
|
| 9 |
+
"ENSG00000129055",
|
| 10 |
+
"ENSG00000112578",
|
| 11 |
+
"ENSG00000088205",
|
| 12 |
+
"ENSG00000166902",
|
| 13 |
+
"ENSG00000171314",
|
| 14 |
+
"ENSG00000124784",
|
| 15 |
+
"ENSG00000075131",
|
| 16 |
+
"ENSG00000134697",
|
| 17 |
+
"ENSG00000153187",
|
| 18 |
+
"ENSG00000089289",
|
| 19 |
+
"ENSG00000143436",
|
| 20 |
+
"ENSG00000101161",
|
| 21 |
+
"ENSG00000101882",
|
| 22 |
+
"ENSG00000196235",
|
| 23 |
+
"ENSG00000108592",
|
| 24 |
+
"ENSG00000070785",
|
| 25 |
+
"ENSG00000169756",
|
| 26 |
+
"ENSG00000090889",
|
| 27 |
+
"ENSG00000139618",
|
| 28 |
+
"ENSG00000100316",
|
| 29 |
+
"ENSG00000204394",
|
| 30 |
+
"ENSG00000198755",
|
| 31 |
+
"ENSG00000167118",
|
| 32 |
+
"ENSG00000107937",
|
| 33 |
+
"ENSG00000183751",
|
| 34 |
+
"ENSG00000141030",
|
| 35 |
+
"ENSG00000094914",
|
| 36 |
+
"ENSG00000126814",
|
| 37 |
+
"ENSG00000004779",
|
| 38 |
+
"ENSG00000067334",
|
| 39 |
+
"ENSG00000131351",
|
| 40 |
+
"ENSG00000204899",
|
| 41 |
+
"ENSG00000149136",
|
| 42 |
+
"ENSG00000183495",
|
| 43 |
+
"ENSG00000175792",
|
| 44 |
+
"ENSG00000188186",
|
| 45 |
+
"ENSG00000111361",
|
| 46 |
+
"ENSG00000166012",
|
| 47 |
+
"ENSG00000173120",
|
| 48 |
+
"ENSG00000129197",
|
| 49 |
+
"ENSG00000196531",
|
| 50 |
+
"ENSG00000173145",
|
| 51 |
+
"ENSG00000184983",
|
| 52 |
+
"ENSG00000103502",
|
| 53 |
+
"ENSG00000127564",
|
| 54 |
+
"ENSG00000132341",
|
| 55 |
+
"ENSG00000175110",
|
| 56 |
+
"ENSG00000151849",
|
| 57 |
+
"ENSG00000125952",
|
| 58 |
+
"ENSG00000126226",
|
| 59 |
+
"ENSG00000177954",
|
| 60 |
+
"ENSG00000172530",
|
| 61 |
+
"ENSG00000239900",
|
| 62 |
+
"ENSG00000173085",
|
| 63 |
+
"ENSG00000185024",
|
| 64 |
+
"ENSG00000183258",
|
| 65 |
+
"ENSG00000213551",
|
| 66 |
+
"ENSG00000197265",
|
| 67 |
+
"ENSG00000136930",
|
| 68 |
+
"ENSG00000127337",
|
| 69 |
+
"ENSG00000156876",
|
| 70 |
+
"ENSG00000119414",
|
| 71 |
+
"ENSG00000130713",
|
| 72 |
+
"ENSG00000121022",
|
| 73 |
+
"ENSG00000164815",
|
| 74 |
+
"ENSG00000173933",
|
| 75 |
+
"ENSG00000168818",
|
| 76 |
+
"ENSG00000116670",
|
| 77 |
+
"ENSG00000162521",
|
| 78 |
+
"ENSG00000117682",
|
| 79 |
+
"ENSG00000169217",
|
| 80 |
+
"ENSG00000134597",
|
| 81 |
+
"ENSG00000197713",
|
| 82 |
+
"ENSG00000113643",
|
| 83 |
+
"ENSG00000062822",
|
| 84 |
+
"ENSG00000158615",
|
| 85 |
+
"ENSG00000119720",
|
| 86 |
+
"ENSG00000173726",
|
| 87 |
+
"ENSG00000134153",
|
| 88 |
+
"ENSG00000142230",
|
| 89 |
+
"ENSG00000143774",
|
| 90 |
+
"ENSG00000160294",
|
| 91 |
+
"ENSG00000092010",
|
| 92 |
+
"ENSG00000108384",
|
| 93 |
+
"ENSG00000136908",
|
| 94 |
+
"ENSG00000142784",
|
| 95 |
+
"ENSG00000108506",
|
| 96 |
+
"ENSG00000135775",
|
| 97 |
+
"ENSG00000196943",
|
| 98 |
+
"ENSG00000172171",
|
| 99 |
+
"ENSG00000159259",
|
| 100 |
+
"ENSG00000126883",
|
| 101 |
+
"ENSG00000125107",
|
| 102 |
+
"ENSG00000150753",
|
| 103 |
+
"ENSG00000132912",
|
| 104 |
+
"ENSG00000130725",
|
| 105 |
+
"ENSG00000063322",
|
| 106 |
+
"ENSG00000137411",
|
| 107 |
+
"ENSG00000108883",
|
| 108 |
+
"ENSG00000115207",
|
| 109 |
+
"ENSG00000163468",
|
| 110 |
+
"ENSG00000162368",
|
| 111 |
+
"ENSG00000146909",
|
| 112 |
+
"ENSG00000136492",
|
| 113 |
+
"ENSG00000090273",
|
| 114 |
+
"ENSG00000073111",
|
| 115 |
+
"ENSG00000161970",
|
| 116 |
+
"ENSG00000021776",
|
| 117 |
+
"ENSG00000157778",
|
| 118 |
+
"ENSG00000099795",
|
| 119 |
+
"ENSG00000106628",
|
| 120 |
+
"ENSG00000101444",
|
| 121 |
+
"ENSG00000110025",
|
| 122 |
+
"ENSG00000108344",
|
| 123 |
+
"ENSG00000123297",
|
| 124 |
+
"ENSG00000167842",
|
| 125 |
+
"ENSG00000163930",
|
| 126 |
+
"ENSG00000127554",
|
| 127 |
+
"ENSG00000167113",
|
| 128 |
+
"ENSG00000184260",
|
| 129 |
+
"ENSG00000116221",
|
| 130 |
+
"ENSG00000044574",
|
| 131 |
+
"ENSG00000008128",
|
| 132 |
+
"ENSG00000168538",
|
| 133 |
+
"ENSG00000163029",
|
| 134 |
+
"ENSG00000114573",
|
| 135 |
+
"ENSG00000106803",
|
| 136 |
+
"ENSG00000088325",
|
| 137 |
+
"ENSG00000135521",
|
| 138 |
+
"ENSG00000093009",
|
| 139 |
+
"ENSG00000136888",
|
| 140 |
+
"ENSG00000163655",
|
| 141 |
+
"ENSG00000143319",
|
| 142 |
+
"ENSG00000104976",
|
| 143 |
+
"ENSG00000132664",
|
| 144 |
+
"ENSG00000102030",
|
| 145 |
+
"ENSG00000157349",
|
| 146 |
+
"ENSG00000174748",
|
| 147 |
+
"ENSG00000112096",
|
| 148 |
+
"ENSG00000278053",
|
| 149 |
+
"ENSG00000138071",
|
| 150 |
+
"ENSG00000100836",
|
| 151 |
+
"ENSG00000155115",
|
| 152 |
+
"ENSG00000090470",
|
| 153 |
+
"ENSG00000080986",
|
| 154 |
+
"ENSG00000133226",
|
| 155 |
+
"ENSG00000108651",
|
| 156 |
+
"ENSG00000131558",
|
| 157 |
+
"ENSG00000165630",
|
| 158 |
+
"ENSG00000119787",
|
| 159 |
+
"ENSG00000196504",
|
| 160 |
+
"ENSG00000159055",
|
| 161 |
+
"ENSG00000159147",
|
| 162 |
+
"ENSG00000196449",
|
| 163 |
+
"ENSG00000164609",
|
| 164 |
+
"ENSG00000091651",
|
| 165 |
+
"ENSG00000014123",
|
| 166 |
+
"ENSG00000136238",
|
| 167 |
+
"ENSG00000115484",
|
| 168 |
+
"ENSG00000198793",
|
| 169 |
+
"ENSG00000211456",
|
| 170 |
+
"ENSG00000249115",
|
| 171 |
+
"ENSG00000186591",
|
| 172 |
+
"ENSG00000061936",
|
| 173 |
+
"ENSG00000183770",
|
| 174 |
+
"ENSG00000144029",
|
| 175 |
+
"ENSG00000119729",
|
| 176 |
+
"ENSG00000184634",
|
| 177 |
+
"ENSG00000136271",
|
| 178 |
+
"ENSG00000100142",
|
| 179 |
+
"ENSG00000008018",
|
| 180 |
+
"ENSG00000103356",
|
| 181 |
+
"ENSG00000124207",
|
| 182 |
+
"ENSG00000108064",
|
| 183 |
+
"ENSG00000176444",
|
| 184 |
+
"ENSG00000083093",
|
| 185 |
+
"ENSG00000140105",
|
| 186 |
+
"ENSG00000010256",
|
| 187 |
+
"ENSG00000127463",
|
| 188 |
+
"ENSG00000167182",
|
| 189 |
+
"ENSG00000139168",
|
| 190 |
+
"ENSG00000174173",
|
| 191 |
+
"ENSG00000119718",
|
| 192 |
+
"ENSG00000239779",
|
| 193 |
+
"ENSG00000130811",
|
| 194 |
+
"ENSG00000149923",
|
| 195 |
+
"ENSG00000088766",
|
| 196 |
+
"ENSG00000152404",
|
| 197 |
+
"ENSG00000131368",
|
| 198 |
+
"ENSG00000231500",
|
| 199 |
+
"ENSG00000115540",
|
| 200 |
+
"ENSG00000163832",
|
| 201 |
+
"ENSG00000113013",
|
| 202 |
+
"ENSG00000141367",
|
| 203 |
+
"ENSG00000134480",
|
| 204 |
+
"ENSG00000100413",
|
| 205 |
+
"ENSG00000010322",
|
| 206 |
+
"ENSG00000181610",
|
| 207 |
+
"ENSG00000112159",
|
| 208 |
+
"ENSG00000168827",
|
| 209 |
+
"ENSG00000101057",
|
| 210 |
+
"ENSG00000149809",
|
| 211 |
+
"ENSG00000170445",
|
| 212 |
+
"ENSG00000107951",
|
| 213 |
+
"ENSG00000137413",
|
| 214 |
+
"ENSG00000198056",
|
| 215 |
+
"ENSG00000243927",
|
| 216 |
+
"ENSG00000105176",
|
| 217 |
+
"ENSG00000135913",
|
| 218 |
+
"ENSG00000180992",
|
| 219 |
+
"ENSG00000072501",
|
| 220 |
+
"ENSG00000142864",
|
| 221 |
+
"ENSG00000165417",
|
| 222 |
+
"ENSG00000188486",
|
| 223 |
+
"ENSG00000060688",
|
| 224 |
+
"ENSG00000213465",
|
| 225 |
+
"ENSG00000104824",
|
| 226 |
+
"ENSG00000172053",
|
| 227 |
+
"ENSG00000112658",
|
| 228 |
+
"ENSG00000070495",
|
| 229 |
+
"ENSG00000161057",
|
| 230 |
+
"ENSG00000164109",
|
| 231 |
+
"ENSG00000165684",
|
| 232 |
+
"ENSG00000176108",
|
| 233 |
+
"ENSG00000141759",
|
| 234 |
+
"ENSG00000204152",
|
| 235 |
+
"ENSG00000165501"
|
| 236 |
+
],
|
| 237 |
+
"shape": [
|
| 238 |
+
233,
|
| 239 |
+
3323
|
| 240 |
+
]
|
| 241 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/data/train_condition_stats.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:213a9cfadf6122065a5f5f5beec43efd18c4ec70015e42e61d142a78fe5c2ae4
|
| 3 |
+
size 54440658
|
tasks/biology.unseen_perturbation_rna_response__seed31415/data/validation_targets.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d40483f6f1f22c35faa091d3150e163655827939fc53366565e948b72a872a72
|
| 3 |
+
size 11697141
|
tasks/biology.unseen_perturbation_rna_response__seed31415/instance_meta.json
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task_id": "biology.unseen_perturbation_rna_response",
|
| 3 |
+
"params_used": {
|
| 4 |
+
"setting": "replogle_supported_seed42",
|
| 5 |
+
"seed": 31415
|
| 6 |
+
},
|
| 7 |
+
"input_files": [
|
| 8 |
+
"data/README.md",
|
| 9 |
+
"data/genes.json",
|
| 10 |
+
"data/manifest.json",
|
| 11 |
+
"data/metadata.json",
|
| 12 |
+
"data/provenance.json",
|
| 13 |
+
"data/test_queries.json",
|
| 14 |
+
"data/train_condition_stats.npz",
|
| 15 |
+
"data/validation_targets.npz"
|
| 16 |
+
],
|
| 17 |
+
"reference_files": [
|
| 18 |
+
"reference.npz",
|
| 19 |
+
"reference.json",
|
| 20 |
+
"results/predictions.npy"
|
| 21 |
+
],
|
| 22 |
+
"generation_time_seconds": 0.9347970616072416
|
| 23 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/prompt_b1.md
ADDED
|
@@ -0,0 +1,441 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Predict downstream RNA responses to unseen gene perturbations
|
| 2 |
+
|
| 3 |
+
Predict the mean measured RNA response for 233 held-out single-gene CRISPR interference conditions in K562 cells. Training observations cover other perturbed genes. The scientific target is the downstream expression response: the directly targeted gene remains in the output, but its weight for its own condition is zero. Return your best complete finite prediction array, including when full credit is not reached. This is prediction from existing measurements, not experimental intervention design.
|
| 4 |
+
|
| 5 |
+
## Work allowance
|
| 6 |
+
|
| 7 |
+
Read `timeout_seconds` in `task_info.json`, record your start time and manage the remaining development allowance. Reasoning, exploration, implementation and verification use that allowance. It is separate from fresh-program replay below. Produce a runnable program and complete output early, retain a working result while improving it, and finish normally within the allowance.
|
| 8 |
+
|
| 9 |
+
## Measurements and fixed split
|
| 10 |
+
|
| 11 |
+
The source is the Replogle K562 essential-gene CRISPRi experiment, sampled six days after perturbation. Every retained matrix feature was matched by its Ensembl ID to original Gene Expression annotation. The released count matrix contains 8,563 verified endogenous RNA features; upstream authors had already restricted their release to genes averaging more than 0.01 UMI per cell. Guide-capture features and other modalities are not expression outputs.
|
| 12 |
+
|
| 13 |
+
For each cell and released RNA gene, the value is
|
| 14 |
+
|
| 15 |
+
`z = ln(1 + 10000 * raw_gene_count / sum_of_all_8563_released_RNA_counts_in_that_cell)`.
|
| 16 |
+
|
| 17 |
+
Normalization precedes output-panel selection. The denominator includes all 8,563 released RNA features, not only the 3,323 selected output features. For each test condition, the target is the gene-wise mean of z over its fixed measured GT-half cells. These are experimental statistics, not predictions produced by an author model.
|
| 18 |
+
|
| 19 |
+
The cohort contains 1,546 canonical target Ensembl IDs with a unique measured RNA identity and at least 64 cells each. The eligible IDs were assigned once to fixed, gene-disjoint training, validation and test sets of 1,082 / 231 / 233 targets. All cells and guides belonging to a target stay in the same split. Training has 188,047 cells including 7,483 controls. Validation and test each have 1,604 control cells.
|
| 20 |
+
|
| 21 |
+
Within each condition, cells are split into two fixed random halves. The first half supplies GT statistics; the other supplies a private technical comparison. The 48 GEM groups and cell halves are technical groupings, not independent biological replicates. The 3,323-feature panel combines 2,000 highly variable RNA genes selected using training cells only with all 1,546 eligible target RNA genes. Every validation and test target therefore occurs as an observed readout column in training data; its held-out perturbation response does not occur there. Target identity is `obs.gene_id` matched to `var.ensembl_id`, rather than a guessed symbol alias.
|
| 22 |
+
|
| 23 |
+
## Public files and schemas
|
| 24 |
+
|
| 25 |
+
All public inputs are in `data/`. Load NPZ files with `allow_pickle=False`. They contain only numeric or Unicode arrays. Preserve all declared row and column orders.
|
| 26 |
+
|
| 27 |
+
- `README.md`: text description of the public files and scientific contract.
|
| 28 |
+
- `train_condition_stats.npz`: `condition_ids` is Unicode[1083], with `control` first and 1,082 training target Ensembl IDs thereafter; `mean_expression` and `variance` are float64[1083,3323]; `n_cells` is int64[1083]. Variance is the per-cell sample variance, with ddof=1, of the normalized log expression. Means and variances use only training cells.
|
| 29 |
+
- `validation_targets.npz`: `condition_ids` is Unicode[231]; `mean_expression` and `metric_weights` are float64[231,3323]; `n_GT` is int64[231]. These measured validation GT-half means and downstream scoring weights are available for tuning and model selection. Technical second-half values are not supplied.
|
| 30 |
+
- `test_queries.json`: `condition_ids` is the ordered list of 233 test target Ensembl strings; `shape` is `[233,3323]`. This file contains identifiers and the requested shape only.
|
| 31 |
+
- `genes.json`: `gene_ids` and `gene_symbols` are ordered lists of 3,323 strings; `raw_column_indices` is the corresponding list of integer indices in the released RNA matrix; `target_id_semantics` is a descriptive string. `gene_ids` defines the output-column order. Symbols are annotations; Ensembl IDs define identity.
|
| 32 |
+
- `metadata.json`: a JSON object defining the dataset, expression scale, split semantics, feature panel, file list, scoring constants and formula, output schema, and resource/data-access contract. Shapes are integer arrays; `scientific_scoring`, `output` and `resource_contract` are explanatory objects.
|
| 33 |
+
- `provenance.json`: dataset citation, license, upstream filtering, feature-ID audit, split summary and public-file hashes. It adds no expression observations. Original full-data gene summary statistics are not supplied as predictive inputs.
|
| 34 |
+
- `manifest.json`: `schema_version` is an integer; `files` maps other public filenames to `sha256` strings and byte counts; `public_test_content` describes the identifier-only test data; `private_evaluation_arrays_included` is false.
|
| 35 |
+
|
| 36 |
+
## Downstream weights and scientific score
|
| 37 |
+
|
| 38 |
+
For each validation or test condition, construct weights from that condition's GT-half cells and the corresponding split's control GT-half cells. Let m, s² and n denote a gene's mean, sample variance and cell count. Compute
|
| 39 |
+
|
| 40 |
+
`t_g = (m_condition,g - m_control,g) / sqrt(s²_condition,g/n_condition + s²_control,g/n_control)`.
|
| 41 |
+
|
| 42 |
+
Set t_g=0 when the denominator is zero. Across all 3,323 genes, min-max scale abs(t_g) to [0,1] and square it. Then set the condition's own canonical target Ensembl gene's entry to exactly zero and renormalize the remaining entries to sum to one. Every frozen condition has positive remaining weight. A degenerate author weight row is an evaluator data error, not a submitted prediction failure. Technical second-half cells never determine these weights.
|
| 43 |
+
|
| 44 |
+
For prediction P, private measured GT means Y and private weights w, the loss is
|
| 45 |
+
|
| 46 |
+
`L = (1/233) * sum_c sum_g w[c,g] * (P[c,g] - Y[c,g])²`.
|
| 47 |
+
|
| 48 |
+
Sum weighted squared errors over genes; do not take an additional gene mean. The score is
|
| 49 |
+
|
| 50 |
+
`score = clip(100 * (B - L) / (B - T), 0, 100)`,
|
| 51 |
+
|
| 52 |
+
`B = 0.03441638519456471`,
|
| 53 |
+
|
| 54 |
+
`T = 0.00714613515111071`.
|
| 55 |
+
|
| 56 |
+
B is the test loss of the equal-weight mean response of the 1,082 noncontrol training conditions. T is the loss of the private technical second-half means against the measured GT-half means, using the same downstream weights. T calibrates sampling reproducibility in this cohort; it is not an independent biological replication result or a proven prediction limit. Losses better than T also receive 100. The evaluator retains raw loss and diagnostics. No method is required to reproduce private experimental labels exactly. Expected range: 100 marks technical half-sample reproducibility, not an attainable target. The strongest method known to the task author scores about 38; scores above about 40 are not expected.
|
| 57 |
+
|
| 58 |
+
## Output and fresh execution
|
| 59 |
+
|
| 60 |
+
Deliver a self-contained `analysis.py` of at most 65,536 bytes (at most 12,288 bytes after zlib level-9 compression, at most 8,192 characters in total across all string, bytes, f-string, numeric and comment tokens, and at most 1,024 characters in any single such token) and `results/predictions.npy` of at most 8,388,608 bytes. The NPY array must have dtype float32 or float64, shape `[233,3323]`, and finite values. Row order is `test_queries.json.condition_ids`; column order is `genes.json.gene_ids`. Include the own-target column even though its condition-specific scoring weight is zero. Finite values of either sign are accepted; there is no extra sign, norm or accuracy gate.
|
| 61 |
+
|
| 62 |
+
`python analysis.py` must recompute your prediction array using `data/` in a fresh workspace. Embed all helper code in that one source file. A source exceeding any size, compression or literal limit above fails the hard gate without being replayed; compute values from `data/` rather than embedding them. Do not depend on a development checkpoint, cached predictions, private labels, or previous scratch files. The hard replay-consistency gate checks both arrays' shape, dtype and finiteness and requires every element to satisfy
|
| 63 |
+
|
| 64 |
+
`abs(submitted - regenerated) <= 1e-5 + 1e-6 * abs(regenerated)`.
|
| 65 |
+
|
| 66 |
+
This is `np.allclose(submitted, regenerated, atol=1e-5, rtol=1e-6)`. It compares your submitted array to your fresh computation, not to experimental truth. Your submission must pass this gate separately from the scientific score.
|
| 67 |
+
|
| 68 |
+
Use only the supplied public data for scientific learning. Training and validation responses may be used with any available computational method or installed library. Do not fetch external datasets, pretrained models, online test responses, or external knowledge resources. Connectivity required by the agent provider is separate from the scientific-data boundary. The same rule applies at all prompt levels.
|
| 69 |
+
|
| 70 |
+
Fresh replay has 1,200 seconds, four assigned CPUs, four numerical-library threads and 8 GiB memory with no swap. Preserve assigned CPU affinity and run offline. The runtime provides Python>=3.11,<3.12, NumPy1.24.3, SciPy1.11.4, pandas1.5.3 and scikit-learn1.3.2 with dependencies. Temporary work belongs in the supplied writable scratch location. Inputs are immutable. The evaluator pre-creates writable `results/predictions.npy`; overwrite it directly using `np.save` on a binary file handle. Do not rename it or create additional result files.
|
| 71 |
+
|
| 72 |
+
A completed low-scoring prediction remains a valid measured attempt. An external agent timeout or evaluator infrastructure failure is not a valid scientific difficulty result. Finish with both the program and completed prediction file.
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
## Complete JSON metadata fields
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
## Additional public JSON field schemas
|
| 79 |
+
|
| 80 |
+
The following lists every field and its type in these three public JSON files. `array[n]` specifies the array length.
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
`data/metadata.json` (object):
|
| 84 |
+
|
| 85 |
+
```text
|
| 86 |
+
"dataset": string
|
| 87 |
+
"expression": string
|
| 88 |
+
"target_semantics": string
|
| 89 |
+
"train_stats": string
|
| 90 |
+
"genes": string
|
| 91 |
+
"validation": string
|
| 92 |
+
"test": string
|
| 93 |
+
"primary_metric": string
|
| 94 |
+
"no_test_response_in_gene_descriptors": boolean
|
| 95 |
+
"files": array[8] of string
|
| 96 |
+
"output_shape": array[2] of integer
|
| 97 |
+
"task_id": string
|
| 98 |
+
"schema_version": integer
|
| 99 |
+
"cell_line": string
|
| 100 |
+
"perturbation_type": string
|
| 101 |
+
"normalization_denominator_features": integer
|
| 102 |
+
"feature_selection": string
|
| 103 |
+
"scientific_scoring": object
|
| 104 |
+
"primary_loss": string
|
| 105 |
+
"own_target_weight": string
|
| 106 |
+
"weight_construction": string
|
| 107 |
+
"score_formula": string
|
| 108 |
+
"B": number
|
| 109 |
+
"T": number
|
| 110 |
+
"B_definition": string
|
| 111 |
+
"T_definition": string
|
| 112 |
+
"GT_origin": string
|
| 113 |
+
"technical_half_limit": string
|
| 114 |
+
"output": object
|
| 115 |
+
"path": string
|
| 116 |
+
"dtype": array[2] of string
|
| 117 |
+
"shape": array[2] of integer
|
| 118 |
+
"finite": boolean
|
| 119 |
+
"maximum_bytes": integer
|
| 120 |
+
"source_path": string
|
| 121 |
+
"maximum_source_bytes": integer
|
| 122 |
+
"maximum_compressed_source_bytes": integer
|
| 123 |
+
"maximum_source_literal_characters": integer
|
| 124 |
+
"maximum_single_literal_characters": integer
|
| 125 |
+
"literal_definition": string
|
| 126 |
+
"resource_contract": object
|
| 127 |
+
"fresh_replay_seconds": integer
|
| 128 |
+
"CPUs": integer
|
| 129 |
+
"threads": integer
|
| 130 |
+
"memory_GiB": integer
|
| 131 |
+
"development_budget": string
|
| 132 |
+
"scientific_data_access": string
|
| 133 |
+
"replay_atol": number
|
| 134 |
+
"replay_rtol": number
|
| 135 |
+
```
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
`data/provenance.json` (object):
|
| 139 |
+
|
| 140 |
+
```text
|
| 141 |
+
"schema_version": integer
|
| 142 |
+
"dataset": string
|
| 143 |
+
"license": string
|
| 144 |
+
"original_paper": string
|
| 145 |
+
"feature_identity_audit": string
|
| 146 |
+
"source_prefilter": string
|
| 147 |
+
"target_identity": string
|
| 148 |
+
"eligible_cohort": string
|
| 149 |
+
"split": string
|
| 150 |
+
"technical_halves": string
|
| 151 |
+
"training_cells": integer
|
| 152 |
+
"training_controls": integer
|
| 153 |
+
"validation_controls": integer
|
| 154 |
+
"test_controls": integer
|
| 155 |
+
"public_numeric_source_hashes": object
|
| 156 |
+
"train_condition_stats.npz": string
|
| 157 |
+
"validation_targets.npz": string
|
| 158 |
+
"offline_boundary": string
|
| 159 |
+
"no_original_full_data_summary_priors": string
|
| 160 |
+
```
|
| 161 |
+
|
| 162 |
+
|
| 163 |
+
`data/manifest.json` (object):
|
| 164 |
+
|
| 165 |
+
```text
|
| 166 |
+
"schema_version": integer
|
| 167 |
+
"files": object
|
| 168 |
+
"README.md": object
|
| 169 |
+
"sha256": string
|
| 170 |
+
"bytes": integer
|
| 171 |
+
"genes.json": object
|
| 172 |
+
"sha256": string
|
| 173 |
+
"bytes": integer
|
| 174 |
+
"metadata.json": object
|
| 175 |
+
"sha256": string
|
| 176 |
+
"bytes": integer
|
| 177 |
+
"provenance.json": object
|
| 178 |
+
"sha256": string
|
| 179 |
+
"bytes": integer
|
| 180 |
+
"test_queries.json": object
|
| 181 |
+
"sha256": string
|
| 182 |
+
"bytes": integer
|
| 183 |
+
"train_condition_stats.npz": object
|
| 184 |
+
"sha256": string
|
| 185 |
+
"bytes": integer
|
| 186 |
+
"validation_targets.npz": object
|
| 187 |
+
"sha256": string
|
| 188 |
+
"bytes": integer
|
| 189 |
+
"public_test_content": string
|
| 190 |
+
"private_evaluation_arrays_included": boolean
|
| 191 |
+
```
|
| 192 |
+
|
| 193 |
+
|
| 194 |
+
## Complete prediction procedure
|
| 195 |
+
|
| 196 |
+
Produce the required analysis.py and results/predictions.npy by implementing and executing the complete prediction procedure below. Preserve its stated transformations, calibration, fold construction, parameter selection and tie breaking. Verify that the program recomputes the finite prediction array from the supplied public inputs and satisfies the common output and replay contract. The procedure is complete; an additional model-search or replacement stage is not required. The prediction target, permitted inputs and scientific score remain as stated above.
|
| 197 |
+
|
| 198 |
+
The eight fixed components are:
|
| 199 |
+
|
| 200 |
+
1. **Original profile RBF.** Apply PCA to the gene-by-training-condition mean-expression matrix after centering each condition column. Retain up to 50 score coordinates, include the control as a zero target descriptor, standardize descriptors using training conditions, and use a centered RBF kernel with inverse median squared-distance bandwidth and ridge penalty 1. This component uses the original training observations only.
|
| 201 |
+
2. **Ordinary response mean.** Average the currently available noncontrol condition means with equal condition weights.
|
| 202 |
+
3. **Weighted response mean.** Reconstruct training downstream weights from the supplied mean, variance and cell-count arrays using the stated Welch-t rule, including the own-target zero. Use the supplied validation weights unchanged. Average each output RNA across available conditions with weights raised to power 0.75; use the original training mean if its weight mass is zero.
|
| 203 |
+
4. **Symbol-family response.** Group target symbols by their leading alphabetic substring. Within a matching group, compute the preceding weighted response mean. Back off to the weighted global response when no matching group or positive weight mass exists. These are string groups made from public annotations, not externally supplied biological labels.
|
| 204 |
+
5. **Character kernel.** Fit case-sensitive character TF-IDF with 3–5 character n-grams and sublinear term frequencies on training-target symbols. Regress condition-response residuals around the original noncontrol training mean using the character inner-product kernel and ridge penalty 0.3. Before transferring a reference response, replace its own-target output by that output's original training mean.
|
| 205 |
+
6. **Incoming-profile neighbors.** Compute each RNA's Welch-t profile across original training interventions. For a training target, replace its own intervention coordinate by the mean of its other coordinates. Apply sign(t)·log1p(abs(t)), center each profile, and normalize it to unit length. Use the 20 most similar known target profiles. Weight neighbors by exp((cosine−largest selected cosine)/0.1), additionally weighting each output by its reference-condition downstream weight raised to 0.35. Replace reference own-target responses as in the character model; use the weighted global response for zero-weight outputs.
|
| 206 |
+
|
| 207 |
+
7. **Symbol-hierarchy kernel.** Build binary namespaced features for the first three characters, the concatenated alphabetic characters, the complete initial ASCII-uppercase run, the first run from the sequence of maximal ASCII-uppercase-letter runs and maximal digit runs, taken separately in encounter order, and the concatenation of the first two runs in that sequence (`[A-Z]+|[0-9]+`, not a mixed alphanumeric run). Preserve case; omit optional run features when absent, and require an initial uppercase run. Fit the dictionary using supplied gene-symbol annotations only and L2-normalize each feature row. Fit their inner-product kernel with ridge0.3 and an unpenalized intercept. Here, replace each reference condition's own-target output by the original control mean, rather than by the original noncontrol training mean. Use exact intercept-corrected leave-one-out predictions before clipping.
|
| 208 |
+
8. **Projected local character neighbors.** Use case-sensitive character1–3 gram raw counts, without IDF or sublinear term frequency, and L2-normalize the count rows. The vocabulary uses supplied gene-symbol annotations only. Select up to30 reference conditions by decreasing cosine similarity, resolving ties by their original row position in the current atlas; exclude a calibration row first. Normalize similarity to the fourth power for the neighbor weights. Average raw measured reference responses without own-target replacement. For zero weight mass, return the original noncontrol training mean. Project the averaged response minus that mean onto the first80 right singular vectors of the original noncontrol training response matrix centered at that mean; retain only positive numerical-rank components, without whitening or output scaling. Restore the same mean. A zero retained rank returns the mean.
|
| 209 |
+
|
| 210 |
+
Numerical descriptors and the output projection use original training observations. The original TF-IDF view uses original training-target symbols; the two new annotation-only views use supplied gene-symbol annotations without validation response labels. For final fitting, the response atlas may include both the training and validation labels, as allowed by the common data contract. Build calibration predictions for every validation row after excluding that row's response from the mean, family and neighbor pools. For the original character model, compute exact leave-one-out predictions from the regularized kernel inverse. For the hierarchy view, let Q=(K+0.3I)^(-1), u=Q1, d=1^T u and M=Q-u u^T/d. With the control-replaced response Y_H, use coefficients M Y_H, intercept (1^T Q Y_H)/d and calibration prediction Y_H[i]-(M Y_H)[i]/M[i,i]. Check finite positive d and relevant M diagonal entries; perform this computation before clipping. The original RBF is already independent of validation labels. Project each base prediction onto nonnegative expression values.
|
| 211 |
+
|
| 212 |
+
## Simplex prior, signed global affine prior, and output-specific ridge
|
| 213 |
+
|
| 214 |
+
First fit the eight-component global nonnegative, sum-to-one mixture on leave-one-out calibration predictions from the current legal response atlas. Its objective is weighted validation error plus `0.01 * max(mean(diag(G_global)), 1e-12) * sum_j (s_j - 1/8)^2`, with predictions and targets centered on the original noncontrol training mean. Denote these simplex coefficients by s. This constraint applies only to that initial prior.
|
| 215 |
+
|
| 216 |
+
For each output RNA g, let `x_cg` be the vector of eight centered base predictions and `t_cg` its centered measured expression. For n calibration conditions, form
|
| 217 |
+
|
| 218 |
+
`G_g = sum_c w_cg * x_cg x_cg^T / n`,
|
| 219 |
+
|
| 220 |
+
`h_g = sum_c w_cg * x_cg * t_cg / n`.
|
| 221 |
+
|
| 222 |
+
Use the original condition-normalized downstream weights; do not renormalize them separately for each output RNA. Sum the per-output statistics to obtain `G_global = sum_g G_g` and `h_global = sum_g h_g`. Fit a shared signed affine prior a with
|
| 223 |
+
|
| 224 |
+
`rho_A = lambda_A * max(trace(G_global)/8, 1e-12)`,
|
| 225 |
+
|
| 226 |
+
`a = solve(G_global + rho_A I, h_global + rho_A s)`.
|
| 227 |
+
|
| 228 |
+
Then fit a separate signed coefficient vector for each output RNA:
|
| 229 |
+
|
| 230 |
+
`rho_g = lambda_G * max(trace(G_g)/8, 1e-12)`,
|
| 231 |
+
|
| 232 |
+
`c_g = solve(G_g + rho_g I, h_g + rho_g a)`.
|
| 233 |
+
|
| 234 |
+
These are two sequential regularized fits, not a jointly optimized hierarchical objective. Neither signed fit is constrained to sum to one or to remain nonnegative. Restore the original training mean after combining centered predictions, then clip negative expression predictions to zero; do not clip the coefficients. There is no additional learned intercept in these three calibration stages; the hierarchy base view has its own unpenalized intercept.
|
| 235 |
+
|
| 236 |
+
The fixed penalty candidates are `lambda_A = [0.01, 0.1, 1.0]` and `lambda_G = [0.1, 1.0, 10.0]`. Enumerate all nine pairs with lambda_A as the outer loop and lambda_G as the inner loop. For the final public-data fit, use three condition folds from `np.array_split(RandomState(20261007).permutation(231), 3)`. For each held-out fold, reconstruct the atlas, leave-one-out calibration bank, simplex prior, global affine fit and per-output fit using only the complementary validation rows plus the original training observations. Held-out labels and weights enter candidate scoring only. Pool condition WMSE across all held-out rows and select the first minimum in the declared pair order. Refit the same functions on all permitted public validation rows using that pair and predict the query IDs in their declared order.
|
| 237 |
+
|
| 238 |
+
## Equations of the eight views and the calibration
|
| 239 |
+
|
| 240 |
+
The equations below define every operator and calibration stage of the procedure above. In this prompt all numerical settings are fixed; they are listed in the next section and replace every "chosen" or "positive" parameter in the equations.
|
| 241 |
+
|
| 242 |
+
### Indices, response pools and fixed origins
|
| 243 |
+
|
| 244 |
+
Let \(T\) be the original noncontrol training conditions, \(V\) the public validation conditions, and \(\mathcal G\) the ordered output RNAs. Each perturbation condition \(i\) has a canonical targeted output column \(j(i)\). Let \(Y_{ig}\) be its measured mean expression and \(W_{ig}\) its downstream weight. For \(i\in T\), reconstruct \(W\) by the common Welch-t/min-max-square rule, including the own-target zero; for \(i\in V\), use the supplied weight row unchanged. Control supplies the original-training descriptor/RBF fit and the hierarchy response replacement described below; it is not a condition in the noncontrol response atlas.
|
| 245 |
+
|
| 246 |
+
The fixed original-training origin is
|
| 247 |
+
|
| 248 |
+
\[
|
| 249 |
+
b_g=|T|^{-1}\sum_{i\in T}Y_{ig}.
|
| 250 |
+
\]
|
| 251 |
+
|
| 252 |
+
For an allowed validation fitting pool \(C\subseteq V\), the response atlas is \(A_C=T\cup C\). A calibration prediction for \(c\in C\) uses the leave-one-condition pool \(A_C\setminus\{c\}\); a query outside the atlas uses all of \(A_C\). Denote the applicable pool by \(A\) in the equations below. Removing a condition means removing its entire response and weight row. For any fixed numerical settings, numerical profile/PCA transforms use original training observations, and the original character TF-IDF vocabulary and inverse-document frequencies use original training-target symbols. The two additional symbol representations below may use all supplied gene-symbol annotations without response labels. Admitting validation responses to an atlas does not refit the original-training transformations or output projection basis.
|
| 253 |
+
|
| 254 |
+
Define a separate cleaned response copy for the original character regression and incoming-profile neighbor transfer only:
|
| 255 |
+
|
| 256 |
+
\[
|
| 257 |
+
\widetilde Y_{ig}=\begin{cases}b_g,&g=j(i),\\Y_{ig},&g\ne j(i).\end{cases}
|
| 258 |
+
\]
|
| 259 |
+
|
| 260 |
+
The RBF fit, ordinary mean and projected local-character view use measured \(Y\). The weighted-global and weighted-family views also use measured \(Y\); their own-target contribution vanishes through \(W_{i,j(i)}=0\). Only the original character and incoming-profile neighbor views use \(\widetilde Y\). The hierarchy view uses its distinct control-replaced response defined below. These column treatments and whole-condition leave-out are separate operations.
|
| 261 |
+
|
| 262 |
+
### The eight response operators
|
| 263 |
+
|
| 264 |
+
**Original-training profile RBF regression.** Let \(T_0\) contain the control and original training conditions. Form the matrix \(M_{gi}=Y_{ig}\), with RNA rows and \(i\in T_0\) columns. Center each condition column across RNA rows and obtain rank-\(r\) PCA scores \(Z=U_r\Sigma_r\). The descriptor of perturbation \(i\) is row \(Z_{j(i),:}\); the control descriptor is zero. Fit coordinate means and standard deviations on these original training-target descriptors plus that zero-control row, and apply the same standardization to every query descriptor. A constant descriptor coordinate has unit divisor.
|
| 265 |
+
|
| 266 |
+
For standardized descriptors \(x\), use a positive bandwidth \(\ell>0\) and positive ridge strength \(\lambda_R>0\), and define \(k_R(x,x')=\exp(-\|x-x'\|^2/\ell^2)\). Let \(K\) be its matrix on \(T_0\), \(H=I-\mathbf1\mathbf1^\top/|T_0|\), and \(K_c=HKH\). The centered query row is
|
| 267 |
+
|
| 268 |
+
\[
|
| 269 |
+
k_c(q,i)=k_R(q,i)-|T_0|^{-1}\sum_{a\in T_0}k_R(q,a)
|
| 270 |
+
-|T_0|^{-1}\sum_{a\in T_0}K_{ai}
|
| 271 |
+
+|T_0|^{-2}\sum_{a,d\in T_0}K_{ad}.
|
| 272 |
+
\]
|
| 273 |
+
|
| 274 |
+
With \(\mu^R=|T_0|^{-1}\sum_{i\in T_0}Y_i\), this view is
|
| 275 |
+
|
| 276 |
+
\[
|
| 277 |
+
B^R_q=\mu^R+k_c(q,:)(K_c+\lambda_R I)^{-1}(Y_{T_0}-\mathbf1(\mu^R)^\top).
|
| 278 |
+
\]
|
| 279 |
+
|
| 280 |
+
The RBF response origin \(\mu^R\) includes control and differs from \(b\). For any fixed numerical settings, this entire view remains fitted on \(T_0\) when validation rows enter another view's atlas.
|
| 281 |
+
|
| 282 |
+
**Ordinary response mean.** For the applicable noncontrol atlas pool,
|
| 283 |
+
|
| 284 |
+
\[
|
| 285 |
+
B^M_{qg}=|A|^{-1}\sum_{i\in A}Y_{ig}.
|
| 286 |
+
\]
|
| 287 |
+
|
| 288 |
+
**Downstream-weighted global mean.** For a positive tempering exponent \(\alpha\),
|
| 289 |
+
|
| 290 |
+
\[
|
| 291 |
+
B^W_{qg}=\frac{\sum_{i\in A}W_{ig}^{\alpha}Y_{ig}}
|
| 292 |
+
{\sum_{i\in A}W_{ig}^{\alpha}}.
|
| 293 |
+
\]
|
| 294 |
+
|
| 295 |
+
A zero denominator falls back to \(b_g\). Weights are indexed by reference condition and output RNA. They are not a query-specific vector of scalar reference weights broadcast over all outputs.
|
| 296 |
+
|
| 297 |
+
**Symbol-family weighted mean.** Let \(f(s)\) be the complete case-preserving initial run of ASCII letters in symbol \(s\), or the entire symbol if there is no initial letter run. Set \(F_q=\{i\in A:f(s_{j(i)})=f(s_{j(q)})\}\). Then
|
| 298 |
+
|
| 299 |
+
\[
|
| 300 |
+
B^F_{qg}=\frac{\sum_{i\in F_q}W_{ig}^{\alpha}Y_{ig}}
|
| 301 |
+
{\sum_{i\in F_q}W_{ig}^{\alpha}}.
|
| 302 |
+
\]
|
| 303 |
+
|
| 304 |
+
An absent family or zero denominator falls back to \(B^W_{qg}\). The same exponent \(\alpha\) is used for global and family means; keys are compared in full, without prefix truncation.
|
| 305 |
+
|
| 306 |
+
**Character inner-product kernel ridge regression.** Fit a case-sensitive character n-gram vocabulary and inverse-document frequencies on original training-target symbols. Use sublinear term frequency and unit-Euclidean-norm TF-IDF vectors \(\phi_i\) for nonzero feature rows; zero feature rows remain zero. Transform query symbols with the same fitted vocabulary. Use a positive ridge strength \(\lambda_C>0\). With \(K^C_{ia}=\phi_i^\top\phi_a\) for \(i,a\in A\) and \(k^C(q,i)=\phi_q^\top\phi_i\),
|
| 307 |
+
|
| 308 |
+
\[
|
| 309 |
+
B^C_q=b+k^C(q,:)(K^C+\lambda_C I)^{-1}
|
| 310 |
+
(\widetilde Y_A-\mathbf1b^\top).
|
| 311 |
+
\]
|
| 312 |
+
|
| 313 |
+
This is a ridge regression operator, not a similarity-normalized mean. For calibration, use the kernel on the atlas with the calibration condition removed. An exact kernel leave-one-out identity or an equivalent refit implements that same exclusion.
|
| 314 |
+
|
| 315 |
+
**Incoming-profile weighted neighbors.** Write \(t_{ig}\) for the signed raw Welch statistic of readout \(g\) under original training intervention \(i\), before the absolute value, min-max scaling, squaring and own-target masking used to construct \(W\). For a training target readout \(g=j(i)\), replace its own-intervention entry by the mean over its other original training interventions. Apply \(u_{gi}=\operatorname{sign}(t_{ig})\log(1+|t_{ig}|)\), center each readout's profile across intervention coordinates, and normalize it to unit Euclidean length. A constant profile becomes zero. Denote the resulting fixed descriptor by \(p_g\).
|
| 316 |
+
|
| 317 |
+
Choose the \(k\) most similar reference targets in \(A\) under \(s_{qi}=p_{j(q)}^\top p_{j(i)}\), calling the set \(N_q\). For positive temperature \(\tau\) and a positive exponent \(\beta\),
|
| 318 |
+
|
| 319 |
+
\[
|
| 320 |
+
a_{qi}=\exp\{(s_{qi}-\max_{d\in N_q}s_{qd})/\tau\},\qquad
|
| 321 |
+
B^N_{qg}=\frac{\sum_{i\in N_q}a_{qi}W_{ig}^{\beta}\widetilde Y_{ig}}
|
| 322 |
+
{\sum_{i\in N_q}a_{qi}W_{ig}^{\beta}}.
|
| 323 |
+
\]
|
| 324 |
+
|
| 325 |
+
A zero denominator falls back to \(B^W_{qg}\). Exclude a calibration condition before neighbor selection. The neighbor exponent \(\beta\) is distinct from the shared global/family exponent \(\alpha\).
|
| 326 |
+
|
| 327 |
+
**Symbol-hierarchy kernel ridge with an unpenalized intercept.** Choose a positive prefix length \(p\). Map each case-preserved target symbol to namespace-separated binary features for: its first \(p\) characters; its alphabetic characters concatenated in their original order; its complete initial ASCII-uppercase-letter run; its first run from the sequence of maximal ASCII-uppercase-letter runs and maximal digit runs, taken separately in encounter order; and the concatenation of the first two runs in that sequence. Alphabetic-character extraction preserves Unicode alphabetic characters; the run extraction separates letter runs from digit runs, equivalent to `[A-Z]+|[0-9]+` rather than `[A-Z0-9]+`. Include the last feature only when two runs exist, and the first-run feature only when a run exists. These feature groups have distinct namespaces even when their string values coincide. Each present dictionary key has value one, rather than a frequency count. The supplied symbols used for this map must have an initial uppercase run.
|
| 328 |
+
|
| 329 |
+
Normalize each nonzero feature row to unit Euclidean length, writing it as \(\psi_i\). The feature dictionary may cover every symbol in the supplied annotations: unused coordinates are zero and no response labels determine it. With a positive ridge strength \(\lambda_H\), define \(K^H_{ia}=\psi_i^\top\psi_a\) on the applicable atlas \(A\). Let \(c^0_g\) denote the original training-control mean and define a distinct response copy
|
| 330 |
+
|
| 331 |
+
\[
|
| 332 |
+
Y^H_{ig}=\begin{cases}c^0_g,&g=j(i),\\Y_{ig},&g\ne j(i).\end{cases}
|
| 333 |
+
\]
|
| 334 |
+
|
| 335 |
+
This replaces an own-target entry with the control mean, not with \(b_g\). Fit kernel ridge with an unpenalized intercept. One equivalent expression is
|
| 336 |
+
|
| 337 |
+
\[
|
| 338 |
+
Q=(K^H+\lambda_H I)^{-1},\quad u=Q\mathbf1,\quad d=\mathbf1^\top u,\quad
|
| 339 |
+
M=Q-uu^\top/d,\quad C_H=MY^H_A,\quad b_H=(\mathbf1^\top QY^H_A)/d,
|
| 340 |
+
\]
|
| 341 |
+
|
| 342 |
+
\[
|
| 343 |
+
B^H_q=b_H+k^H(q,:)C_H,\qquad k^H(q,i)=\psi_q^\top\psi_i.
|
| 344 |
+
\]
|
| 345 |
+
|
| 346 |
+
For calibration, remove the condition before fitting or use the equivalent intercept-corrected leave-one-out identity on the full legal atlas:
|
| 347 |
+
|
| 348 |
+
\[
|
| 349 |
+
B^{H,\mathrm{LOO}}_i=Y^H_i-(C_H)_i/M_{ii}.
|
| 350 |
+
\]
|
| 351 |
+
|
| 352 |
+
Apply this identity to the raw response prediction before nonnegative clipping. Require positive finite denominators; a zero query kernel follows the fitted intercept, rather than an unrelated mean fallback. This operator differs from the preceding fixed-origin TF-IDF ridge in both its annotation features and its unpenalized intercept.
|
| 353 |
+
|
| 354 |
+
**Projected local character-count neighbors.** Form case-sensitive character n-gram count vectors over a selected positive n-gram-length range, counting repeated substrings. Use ordinary term counts with neither inverse-document-frequency weighting nor sublinear term-frequency transformation, then normalize nonzero vectors to unit Euclidean length; zero rows remain zero. The annotation-only vocabulary may include every supplied gene symbol. Let these vectors be \(\chi_i\) and let \(r_{qi}=\chi_q^\top\chi_i\), a nonnegative cosine similarity.
|
| 355 |
+
|
| 356 |
+
Choose up to \(k_L\) nearest reference conditions in the applicable atlas, excluding a calibration condition before selection. Fix a deterministic tie rule. With positive power \(\eta\), define
|
| 357 |
+
|
| 358 |
+
\[
|
| 359 |
+
\omega_{qi}=\frac{r_{qi}^{\eta}}{\sum_{a\in N^L_q}r_{qa}^{\eta}},\qquad
|
| 360 |
+
\overline Y^L_q=\sum_{i\in N^L_q}\omega_{qi}Y_i.
|
| 361 |
+
\]
|
| 362 |
+
|
| 363 |
+
These reference responses are measured \(Y\), with no own-target replacement and no downstream-weight multiplier. Zero similarity-weight mass falls back to \(b\). From the original noncontrol training matrix \(Y_T-\mathbf1b^\top\), compute orthonormal right singular vectors and retain a chosen rank \(r_L\) within its numerical rank, collected as rows of \(V_L\). Do not standardize output columns, whiten the basis or multiply it by singular values. The view is
|
| 364 |
+
|
| 365 |
+
\[
|
| 366 |
+
B^L_q=b+(\overline Y^L_q-b)V_L^\top V_L.
|
| 367 |
+
\]
|
| 368 |
+
|
| 369 |
+
A zero retained rank returns \(b\). This output projection is fitted only on original training responses and remains fixed when validation rows enter or leave the atlas. It differs from incoming-profile neighbors: neighbor identity comes from character counts, while the training-only projection regularizes the transferred output response.
|
| 370 |
+
|
| 371 |
+
Project each of the eight base expression predictions onto nonnegative values before calibration. Numerical zero-mass and zero-norm tolerances remain implementation choices; preserve the stated fallbacks. Below, \(B^{(a)}\) denotes these projected base predictions, with one fixed ordering of the eight components.
|
| 372 |
+
|
| 373 |
+
### Three sequential calibration fits
|
| 374 |
+
|
| 375 |
+
Build leave-one-condition calibration predictions for every \(c\in C\) using its applicable atlas exclusion. Set the eight-component vector and centered target to
|
| 376 |
+
|
| 377 |
+
\[
|
| 378 |
+
X_{cg}=(B^{(a)}_{cg}-b_g)_{a=1}^{8},\qquad z_{cg}=Y_{cg}-b_g.
|
| 379 |
+
\]
|
| 380 |
+
|
| 381 |
+
All calibration observations and their weights come from \(C\). Define
|
| 382 |
+
|
| 383 |
+
\[
|
| 384 |
+
G_g=|C|^{-1}\sum_{c\in C}W_{cg}X_{cg}X_{cg}^\top,\quad
|
| 385 |
+
h_g=|C|^{-1}\sum_{c\in C}W_{cg}X_{cg}z_{cg},\quad
|
| 386 |
+
G=\sum_g G_g,\quad h=\sum_g h_g.
|
| 387 |
+
\]
|
| 388 |
+
|
| 389 |
+
Let \(s=\max(\operatorname{tr}(G)/8,\epsilon)\), \(s_g=\max(\operatorname{tr}(G_g)/8,\epsilon)\), with a positive numerical floor \(\epsilon\), and let \(v_0=\mathbf1/8\). Fit, in this order,
|
| 390 |
+
|
| 391 |
+
\[
|
| 392 |
+
v_S=\arg\min_{v\ge0,\;\mathbf1^\top v=1}
|
| 393 |
+
\{v^\top Gv-2h^\top v+\lambda_Ss\|v-v_0\|^2\},
|
| 394 |
+
\]
|
| 395 |
+
|
| 396 |
+
\[
|
| 397 |
+
v_A=\arg\min_{v\in\mathbb R^8}
|
| 398 |
+
\{v^\top Gv-2h^\top v+\lambda_As\|v-v_S\|^2\},
|
| 399 |
+
\]
|
| 400 |
+
|
| 401 |
+
\[
|
| 402 |
+
v_g=\arg\min_{v\in\mathbb R^8}
|
| 403 |
+
\{v^\top G_gv-2h_g^\top v+\lambda_Gs_g\|v-v_A\|^2\}.
|
| 404 |
+
\]
|
| 405 |
+
|
| 406 |
+
The Gram forms express the corresponding condition/output-weighted squared losses, up to constants independent of the coefficients. Only \(v_S\) lies on a simplex. The global signed vector \(v_A\) and output-specific signed vectors \(v_g\) have neither a nonnegativity nor a sum-to-one constraint. The positive strengths \(\lambda_S,\lambda_A,\lambda_G\) are numerical choices; \(\lambda_G\) is a shared scalar with output-specific scales \(s_g\). These are sequential fits toward already fitted priors, not one joint optimization over all stages. The fixed origin supplies the response offset; there is no additional fitted intercept in these equations.
|
| 407 |
+
|
| 408 |
+
For a final query, the delivered expression is
|
| 409 |
+
|
| 410 |
+
\[
|
| 411 |
+
\widehat Y_{qg}=\max\{0,\ b_g+v_g^\top(B^{(a)}_{qg}-b_g)_{a=1}^{8}\}.
|
| 412 |
+
\]
|
| 413 |
+
|
| 414 |
+
This equation defines the final combination of the eight views. Use the fixed numerical settings below and return the predictions in the common required order.
|
| 415 |
+
|
| 416 |
+
## Fixed numerical settings and implementation details
|
| 417 |
+
|
| 418 |
+
- **Row orders.** The training atlas rows are the 1,082 noncontrol rows of `train_condition_stats.npz` in file order. A validation pool appends its validation rows in `validation_targets.npz` order. Output columns follow `genes.json.gene_ids`; condition \(i\) maps to the column whose Ensembl ID equals its condition ID.
|
| 419 |
+
- **Training weights and Welch profiles.** For every training row, \(t_{ig}=(m_{ig}-m_{0g})/\sqrt{v_{ig}/n_i+v_{0g}/n_0}\) with the control row 0, and \(t=0\) for a zero denominator. The weight row is the min-max-scaled \(|t|\) squared, with its own-target entry zero and the row renormalized to sum to one.
|
| 420 |
+
- **Profile RBF.** The PCA matrix holds the 3,323 RNAs as rows and all 1,083 training rows (control first) as columns. Each column is centered over RNAs and \(r=\min(50,\text{numerical capacity})\) scores \(U_r\Sigma_r\) are kept. The 1,083 descriptors are the zero control row and the training targets' score rows. They are standardized with their mean and population standard deviation (ddof 0; a zero deviation is replaced by 1). Then \(\ell^2\) is the median of the positive pairwise squared distances among these descriptors (upper triangle), and \(\lambda_R=1\). The view is fitted on all 1,083 rows, with \(\mu^R\) their mean.
|
| 421 |
+
- **Weighted global and family means.** \(\alpha=0.75\). A weight mass at or below \(10^{-30}\) uses the stated fallback. The family key is the match of `[A-Za-z]+` at the start of the symbol, or the whole symbol when there is none.
|
| 422 |
+
- **Character ridge.** Use a case-sensitive character TF-IDF over 3–5-grams, with smoothed inverse document frequency \(\log((1+N)/(1+\mathrm{df}))+1\), sublinear term frequency \(1+\log(\mathrm{tf})\) and L2 row normalization. It is fitted on the 1,082 training-target symbols and \(\lambda_C=0.3\). Leave-one-out calibration uses \(\widehat Y_i=b+(\widetilde Y_i-b)-[(K^C+\lambda_CI)^{-1}(\widetilde Y_A-\mathbf1b^\top)]_i/[(K^C+\lambda_CI)^{-1}]_{ii}\).
|
| 423 |
+
- **Incoming-profile neighbors.** Profiles have one coordinate per training intervention. Each centered profile is divided by \(\max(\|\cdot\|,10^{-30})\). Use \(k=20\) neighbors with the largest similarity, \(\tau=0.1\), \(\beta=0.35\) and a mass floor of \(10^{-30}\).
|
| 424 |
+
- **Symbol hierarchy.** \(p=3\) and \(\lambda_H=0.3\), with the five namespaced binary keys built over all 3,323 supplied symbols and L2-normalized.
|
| 425 |
+
- **Projected local character neighbors.** Use raw character 1–3-gram counts over all supplied symbols, L2-normalized. Take up to \(k_L=30\) references in descending similarity, breaking ties by atlas row order, with \(\eta=4\). The basis is the leading \(r_L=\min(80,\text{numerical rank})\) right singular vectors of \(Y_T-\mathbf1b^\top\). The numerical rank counts singular values above \(\max(\text{shape})\cdot\epsilon_{\text{machine}}\cdot\sigma_1\).
|
| 426 |
+
- **Calibration.** Order the components as: RBF, ordinary mean, weighted global, family, character ridge, incoming neighbors, hierarchy, local character. Symmetrize each \(G_g\) as \((G_g+G_g^\top)/2\), with \(\epsilon=10^{-12}\). For the simplex prior, use \(\lambda_S=0.01\) and \(\operatorname{tr}(G)/8\) as its scale. Solve it from the uniform start under nonnegativity and sum-to-one constraints, to a tolerance near \(10^{-12}\). Then \(v_A\) and \(v_g\) are the closed-form ridge solutions \((G+\lambda_AsI)^{-1}(h+\lambda_Asv_S)\) and \((G_g+\lambda_Gs_gI)^{-1}(h_g+\lambda_Gs_gv_A)\).
|
| 427 |
+
- **Selection and final fit.** Candidates are \(\lambda_A\in\{0.01,0.1,1\}\) (outer loop) and \(\lambda_G\in\{0.1,1,10\}\) (inner loop). The three folds are `np.array_split(np.random.RandomState(20261007).permutation(231), 3)` over validation row positions. For each fold, fit the atlas, calibration bank and coefficients on the other two folds. Predict the held-out rows as queries outside the atlas and record each row's weighted squared error. Average over all 231 rows and choose the first minimum. Refit on all 231 validation rows with the chosen pair, predict `test_queries.json.condition_ids` in order, and save float64 predictions.
|
| 428 |
+
|
| 429 |
+
## Consistent validation and final fitting
|
| 430 |
+
|
| 431 |
+
The following validation guidance applies if you assess an optional change to the specified procedure. Additional model search or a new outer-validation experiment is not required to complete the prescribed prediction procedure and its common output/replay checks.
|
| 432 |
+
|
| 433 |
+
Use a single reusable prediction implementation for both held-out validation conditions and final query conditions. Change only which response rows are available to that fit. Do not validate a simplified surrogate and then silently change family weights, normalization, masking, atlas membership, or residual corrections in the final program.
|
| 434 |
+
|
| 435 |
+
Create one condition-ID-to-fold manifest and reuse it for every component. Store each out-of-fold prediction together with its condition IDs, output gene IDs, and fold assignments. Align arrays by these identifiers and assert identical row order, column order and fold assignments before combining component prediction banks. Equal array shapes are insufficient. If the folds differ, regenerate the affected predictions under the shared manifest before comparing or fitting a mixture.
|
| 436 |
+
|
| 437 |
+
Evaluate the full fitted ensemble with nested validation. For each outer held-out fold, fit every label-dependent component, mixture coefficient, routing choice, shrinkage parameter and residual offset using only the other folds. Inner validation can choose settings within that outer training pool. Then evaluate the resulting fixed procedure on the outer fold. Fitting ensemble coefficients to a bank of out-of-fold predictions and reporting loss on those same rows is a calibration objective, not an untouched estimate of the fitted ensemble's generalization error.
|
| 438 |
+
|
| 439 |
+
Compare the supplied procedure and any proposed change on exactly the same held-out conditions and weights. Once a procedure is chosen using public data, refit it on the permitted training and validation responses, retaining the same prediction functions for the ordered query IDs. Fresh-source replay verifies reproducibility; it does not verify that a validation procedure was correctly isolated.
|
| 440 |
+
|
| 441 |
+
For example, if an alternative adds a residual bias or offset, let F be an outer held-out fold and T the remaining validation conditions. Fit that correction from inner out-of-fold residuals rebuilt entirely inside T, with F excluded from every inner fit. Do not build a residual bank using all validation conditions and then merely remove F's rows: fits behind the other rows may already have used F's labels. After method selection, fitting the final procedure on all permitted public validation conditions remains allowed.
|
tasks/biology.unseen_perturbation_rna_response__seed31415/prompt_b2.md
ADDED
|
@@ -0,0 +1,379 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Predict downstream RNA responses to unseen gene perturbations
|
| 2 |
+
|
| 3 |
+
Predict the mean measured RNA response for 233 held-out single-gene CRISPR interference conditions in K562 cells. Training observations cover other perturbed genes. The scientific target is the downstream expression response: the directly targeted gene remains in the output, but its weight for its own condition is zero. Return your best complete finite prediction array, including when full credit is not reached. This is prediction from existing measurements, not experimental intervention design.
|
| 4 |
+
|
| 5 |
+
## Work allowance
|
| 6 |
+
|
| 7 |
+
Read `timeout_seconds` in `task_info.json`, record your start time and manage the remaining development allowance. Reasoning, exploration, implementation and verification use that allowance. It is separate from fresh-program replay below. Produce a runnable program and complete output early, retain a working result while improving it, and finish normally within the allowance.
|
| 8 |
+
|
| 9 |
+
## Measurements and fixed split
|
| 10 |
+
|
| 11 |
+
The source is the Replogle K562 essential-gene CRISPRi experiment, sampled six days after perturbation. Every retained matrix feature was matched by its Ensembl ID to original Gene Expression annotation. The released count matrix contains 8,563 verified endogenous RNA features; upstream authors had already restricted their release to genes averaging more than 0.01 UMI per cell. Guide-capture features and other modalities are not expression outputs.
|
| 12 |
+
|
| 13 |
+
For each cell and released RNA gene, the value is
|
| 14 |
+
|
| 15 |
+
`z = ln(1 + 10000 * raw_gene_count / sum_of_all_8563_released_RNA_counts_in_that_cell)`.
|
| 16 |
+
|
| 17 |
+
Normalization precedes output-panel selection. The denominator includes all 8,563 released RNA features, not only the 3,323 selected output features. For each test condition, the target is the gene-wise mean of z over its fixed measured GT-half cells. These are experimental statistics, not predictions produced by an author model.
|
| 18 |
+
|
| 19 |
+
The cohort contains 1,546 canonical target Ensembl IDs with a unique measured RNA identity and at least 64 cells each. The eligible IDs were assigned once to fixed, gene-disjoint training, validation and test sets of 1,082 / 231 / 233 targets. All cells and guides belonging to a target stay in the same split. Training has 188,047 cells including 7,483 controls. Validation and test each have 1,604 control cells.
|
| 20 |
+
|
| 21 |
+
Within each condition, cells are split into two fixed random halves. The first half supplies GT statistics; the other supplies a private technical comparison. The 48 GEM groups and cell halves are technical groupings, not independent biological replicates. The 3,323-feature panel combines 2,000 highly variable RNA genes selected using training cells only with all 1,546 eligible target RNA genes. Every validation and test target therefore occurs as an observed readout column in training data; its held-out perturbation response does not occur there. Target identity is `obs.gene_id` matched to `var.ensembl_id`, rather than a guessed symbol alias.
|
| 22 |
+
|
| 23 |
+
## Public files and schemas
|
| 24 |
+
|
| 25 |
+
All public inputs are in `data/`. Load NPZ files with `allow_pickle=False`. They contain only numeric or Unicode arrays. Preserve all declared row and column orders.
|
| 26 |
+
|
| 27 |
+
- `README.md`: text description of the public files and scientific contract.
|
| 28 |
+
- `train_condition_stats.npz`: `condition_ids` is Unicode[1083], with `control` first and 1,082 training target Ensembl IDs thereafter; `mean_expression` and `variance` are float64[1083,3323]; `n_cells` is int64[1083]. Variance is the per-cell sample variance, with ddof=1, of the normalized log expression. Means and variances use only training cells.
|
| 29 |
+
- `validation_targets.npz`: `condition_ids` is Unicode[231]; `mean_expression` and `metric_weights` are float64[231,3323]; `n_GT` is int64[231]. These measured validation GT-half means and downstream scoring weights are available for tuning and model selection. Technical second-half values are not supplied.
|
| 30 |
+
- `test_queries.json`: `condition_ids` is the ordered list of 233 test target Ensembl strings; `shape` is `[233,3323]`. This file contains identifiers and the requested shape only.
|
| 31 |
+
- `genes.json`: `gene_ids` and `gene_symbols` are ordered lists of 3,323 strings; `raw_column_indices` is the corresponding list of integer indices in the released RNA matrix; `target_id_semantics` is a descriptive string. `gene_ids` defines the output-column order. Symbols are annotations; Ensembl IDs define identity.
|
| 32 |
+
- `metadata.json`: a JSON object defining the dataset, expression scale, split semantics, feature panel, file list, scoring constants and formula, output schema, and resource/data-access contract. Shapes are integer arrays; `scientific_scoring`, `output` and `resource_contract` are explanatory objects.
|
| 33 |
+
- `provenance.json`: dataset citation, license, upstream filtering, feature-ID audit, split summary and public-file hashes. It adds no expression observations. Original full-data gene summary statistics are not supplied as predictive inputs.
|
| 34 |
+
- `manifest.json`: `schema_version` is an integer; `files` maps other public filenames to `sha256` strings and byte counts; `public_test_content` describes the identifier-only test data; `private_evaluation_arrays_included` is false.
|
| 35 |
+
|
| 36 |
+
## Downstream weights and scientific score
|
| 37 |
+
|
| 38 |
+
For each validation or test condition, construct weights from that condition's GT-half cells and the corresponding split's control GT-half cells. Let m, s² and n denote a gene's mean, sample variance and cell count. Compute
|
| 39 |
+
|
| 40 |
+
`t_g = (m_condition,g - m_control,g) / sqrt(s²_condition,g/n_condition + s²_control,g/n_control)`.
|
| 41 |
+
|
| 42 |
+
Set t_g=0 when the denominator is zero. Across all 3,323 genes, min-max scale abs(t_g) to [0,1] and square it. Then set the condition's own canonical target Ensembl gene's entry to exactly zero and renormalize the remaining entries to sum to one. Every frozen condition has positive remaining weight. A degenerate author weight row is an evaluator data error, not a submitted prediction failure. Technical second-half cells never determine these weights.
|
| 43 |
+
|
| 44 |
+
For prediction P, private measured GT means Y and private weights w, the loss is
|
| 45 |
+
|
| 46 |
+
`L = (1/233) * sum_c sum_g w[c,g] * (P[c,g] - Y[c,g])²`.
|
| 47 |
+
|
| 48 |
+
Sum weighted squared errors over genes; do not take an additional gene mean. The score is
|
| 49 |
+
|
| 50 |
+
`score = clip(100 * (B - L) / (B - T), 0, 100)`,
|
| 51 |
+
|
| 52 |
+
`B = 0.03441638519456471`,
|
| 53 |
+
|
| 54 |
+
`T = 0.00714613515111071`.
|
| 55 |
+
|
| 56 |
+
B is the test loss of the equal-weight mean response of the 1,082 noncontrol training conditions. T is the loss of the private technical second-half means against the measured GT-half means, using the same downstream weights. T calibrates sampling reproducibility in this cohort; it is not an independent biological replication result or a proven prediction limit. Losses better than T also receive 100. The evaluator retains raw loss and diagnostics. No method is required to reproduce private experimental labels exactly. Expected range: 100 marks technical half-sample reproducibility, not an attainable target. The strongest method known to the task author scores about 38; scores above about 40 are not expected.
|
| 57 |
+
|
| 58 |
+
## Output and fresh execution
|
| 59 |
+
|
| 60 |
+
Deliver a self-contained `analysis.py` of at most 65,536 bytes (at most 12,288 bytes after zlib level-9 compression, at most 8,192 characters in total across all string, bytes, f-string, numeric and comment tokens, and at most 1,024 characters in any single such token) and `results/predictions.npy` of at most 8,388,608 bytes. The NPY array must have dtype float32 or float64, shape `[233,3323]`, and finite values. Row order is `test_queries.json.condition_ids`; column order is `genes.json.gene_ids`. Include the own-target column even though its condition-specific scoring weight is zero. Finite values of either sign are accepted; there is no extra sign, norm or accuracy gate.
|
| 61 |
+
|
| 62 |
+
`python analysis.py` must recompute your prediction array using `data/` in a fresh workspace. Embed all helper code in that one source file. A source exceeding any size, compression or literal limit above fails the hard gate without being replayed; compute values from `data/` rather than embedding them. Do not depend on a development checkpoint, cached predictions, private labels, or previous scratch files. The hard replay-consistency gate checks both arrays' shape, dtype and finiteness and requires every element to satisfy
|
| 63 |
+
|
| 64 |
+
`abs(submitted - regenerated) <= 1e-5 + 1e-6 * abs(regenerated)`.
|
| 65 |
+
|
| 66 |
+
This is `np.allclose(submitted, regenerated, atol=1e-5, rtol=1e-6)`. It compares your submitted array to your fresh computation, not to experimental truth. Your submission must pass this gate separately from the scientific score.
|
| 67 |
+
|
| 68 |
+
Use only the supplied public data for scientific learning. Training and validation responses may be used with any available computational method or installed library. Do not fetch external datasets, pretrained models, online test responses, or external knowledge resources. Connectivity required by the agent provider is separate from the scientific-data boundary. The same rule applies at all prompt levels.
|
| 69 |
+
|
| 70 |
+
Fresh replay has 1,200 seconds, four assigned CPUs, four numerical-library threads and 8 GiB memory with no swap. Preserve assigned CPU affinity and run offline. The runtime provides Python>=3.11,<3.12, NumPy1.24.3, SciPy1.11.4, pandas1.5.3 and scikit-learn1.3.2 with dependencies. Temporary work belongs in the supplied writable scratch location. Inputs are immutable. The evaluator pre-creates writable `results/predictions.npy`; overwrite it directly using `np.save` on a binary file handle. Do not rename it or create additional result files.
|
| 71 |
+
|
| 72 |
+
A completed low-scoring prediction remains a valid measured attempt. An external agent timeout or evaluator infrastructure failure is not a valid scientific difficulty result. Finish with both the program and completed prediction file.
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
## Complete JSON metadata fields
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
## Additional public JSON field schemas
|
| 79 |
+
|
| 80 |
+
The following lists every field and its type in these three public JSON files. `array[n]` specifies the array length.
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
`data/metadata.json` (object):
|
| 84 |
+
|
| 85 |
+
```text
|
| 86 |
+
"dataset": string
|
| 87 |
+
"expression": string
|
| 88 |
+
"target_semantics": string
|
| 89 |
+
"train_stats": string
|
| 90 |
+
"genes": string
|
| 91 |
+
"validation": string
|
| 92 |
+
"test": string
|
| 93 |
+
"primary_metric": string
|
| 94 |
+
"no_test_response_in_gene_descriptors": boolean
|
| 95 |
+
"files": array[8] of string
|
| 96 |
+
"output_shape": array[2] of integer
|
| 97 |
+
"task_id": string
|
| 98 |
+
"schema_version": integer
|
| 99 |
+
"cell_line": string
|
| 100 |
+
"perturbation_type": string
|
| 101 |
+
"normalization_denominator_features": integer
|
| 102 |
+
"feature_selection": string
|
| 103 |
+
"scientific_scoring": object
|
| 104 |
+
"primary_loss": string
|
| 105 |
+
"own_target_weight": string
|
| 106 |
+
"weight_construction": string
|
| 107 |
+
"score_formula": string
|
| 108 |
+
"B": number
|
| 109 |
+
"T": number
|
| 110 |
+
"B_definition": string
|
| 111 |
+
"T_definition": string
|
| 112 |
+
"GT_origin": string
|
| 113 |
+
"technical_half_limit": string
|
| 114 |
+
"output": object
|
| 115 |
+
"path": string
|
| 116 |
+
"dtype": array[2] of string
|
| 117 |
+
"shape": array[2] of integer
|
| 118 |
+
"finite": boolean
|
| 119 |
+
"maximum_bytes": integer
|
| 120 |
+
"source_path": string
|
| 121 |
+
"maximum_source_bytes": integer
|
| 122 |
+
"maximum_compressed_source_bytes": integer
|
| 123 |
+
"maximum_source_literal_characters": integer
|
| 124 |
+
"maximum_single_literal_characters": integer
|
| 125 |
+
"literal_definition": string
|
| 126 |
+
"resource_contract": object
|
| 127 |
+
"fresh_replay_seconds": integer
|
| 128 |
+
"CPUs": integer
|
| 129 |
+
"threads": integer
|
| 130 |
+
"memory_GiB": integer
|
| 131 |
+
"development_budget": string
|
| 132 |
+
"scientific_data_access": string
|
| 133 |
+
"replay_atol": number
|
| 134 |
+
"replay_rtol": number
|
| 135 |
+
```
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
`data/provenance.json` (object):
|
| 139 |
+
|
| 140 |
+
```text
|
| 141 |
+
"schema_version": integer
|
| 142 |
+
"dataset": string
|
| 143 |
+
"license": string
|
| 144 |
+
"original_paper": string
|
| 145 |
+
"feature_identity_audit": string
|
| 146 |
+
"source_prefilter": string
|
| 147 |
+
"target_identity": string
|
| 148 |
+
"eligible_cohort": string
|
| 149 |
+
"split": string
|
| 150 |
+
"technical_halves": string
|
| 151 |
+
"training_cells": integer
|
| 152 |
+
"training_controls": integer
|
| 153 |
+
"validation_controls": integer
|
| 154 |
+
"test_controls": integer
|
| 155 |
+
"public_numeric_source_hashes": object
|
| 156 |
+
"train_condition_stats.npz": string
|
| 157 |
+
"validation_targets.npz": string
|
| 158 |
+
"offline_boundary": string
|
| 159 |
+
"no_original_full_data_summary_priors": string
|
| 160 |
+
```
|
| 161 |
+
|
| 162 |
+
|
| 163 |
+
`data/manifest.json` (object):
|
| 164 |
+
|
| 165 |
+
```text
|
| 166 |
+
"schema_version": integer
|
| 167 |
+
"files": object
|
| 168 |
+
"README.md": object
|
| 169 |
+
"sha256": string
|
| 170 |
+
"bytes": integer
|
| 171 |
+
"genes.json": object
|
| 172 |
+
"sha256": string
|
| 173 |
+
"bytes": integer
|
| 174 |
+
"metadata.json": object
|
| 175 |
+
"sha256": string
|
| 176 |
+
"bytes": integer
|
| 177 |
+
"provenance.json": object
|
| 178 |
+
"sha256": string
|
| 179 |
+
"bytes": integer
|
| 180 |
+
"test_queries.json": object
|
| 181 |
+
"sha256": string
|
| 182 |
+
"bytes": integer
|
| 183 |
+
"train_condition_stats.npz": object
|
| 184 |
+
"sha256": string
|
| 185 |
+
"bytes": integer
|
| 186 |
+
"validation_targets.npz": object
|
| 187 |
+
"sha256": string
|
| 188 |
+
"bytes": integer
|
| 189 |
+
"public_test_content": string
|
| 190 |
+
"private_evaluation_arrays_included": boolean
|
| 191 |
+
```
|
| 192 |
+
|
| 193 |
+
## Guided eight-view estimator
|
| 194 |
+
|
| 195 |
+
Implement, validate and deliver the eight-view estimator defined below. Its final prediction is the specified sequentially calibrated combination of these eight operators. Choose its numerical settings using the supplied public validation data, then execute the same estimator for the ordered queries. The equations specify the method; descriptor dimensions, bandwidths, regularization strengths, neighbor settings, candidate sets, fold construction and numerical implementation remain your choices. The common artifact, resource, scientific-score and replay contract above is unchanged. Algorithm identity is not an additional scoring gate.
|
| 196 |
+
|
| 197 |
+
### Indices, response pools and fixed origins
|
| 198 |
+
|
| 199 |
+
Let \(T\) be the original noncontrol training conditions, \(V\) the public validation conditions, and \(\mathcal G\) the ordered output RNAs. Each perturbation condition \(i\) has a canonical targeted output column \(j(i)\). Let \(Y_{ig}\) be its measured mean expression and \(W_{ig}\) its downstream weight. For \(i\in T\), reconstruct \(W\) by the common Welch-t/min-max-square rule, including the own-target zero; for \(i\in V\), use the supplied weight row unchanged. Control supplies the original-training descriptor/RBF fit and the hierarchy response replacement described below; it is not a condition in the noncontrol response atlas.
|
| 200 |
+
|
| 201 |
+
The fixed original-training origin is
|
| 202 |
+
|
| 203 |
+
\[
|
| 204 |
+
b_g=|T|^{-1}\sum_{i\in T}Y_{ig}.
|
| 205 |
+
\]
|
| 206 |
+
|
| 207 |
+
For an allowed validation fitting pool \(C\subseteq V\), the response atlas is \(A_C=T\cup C\). A calibration prediction for \(c\in C\) uses the leave-one-condition pool \(A_C\setminus\{c\}\); a query outside the atlas uses all of \(A_C\). Denote the applicable pool by \(A\) in the equations below. Removing a condition means removing its entire response and weight row. For any fixed numerical settings, numerical profile/PCA transforms use original training observations, and the original character TF-IDF vocabulary and inverse-document frequencies use original training-target symbols. The two additional symbol representations below may use all supplied gene-symbol annotations without response labels. Admitting validation responses to an atlas does not refit the original-training transformations or output projection basis.
|
| 208 |
+
|
| 209 |
+
Define a separate cleaned response copy for the original character regression and incoming-profile neighbor transfer only:
|
| 210 |
+
|
| 211 |
+
\[
|
| 212 |
+
\widetilde Y_{ig}=\begin{cases}b_g,&g=j(i),\\Y_{ig},&g\ne j(i).\end{cases}
|
| 213 |
+
\]
|
| 214 |
+
|
| 215 |
+
The RBF fit, ordinary mean and projected local-character view use measured \(Y\). The weighted-global and weighted-family views also use measured \(Y\); their own-target contribution vanishes through \(W_{i,j(i)}=0\). Only the original character and incoming-profile neighbor views use \(\widetilde Y\). The hierarchy view uses its distinct control-replaced response defined below. These column treatments and whole-condition leave-out are separate operations.
|
| 216 |
+
|
| 217 |
+
### The eight response operators
|
| 218 |
+
|
| 219 |
+
**Original-training profile RBF regression.** Let \(T_0\) contain the control and original training conditions. Form the matrix \(M_{gi}=Y_{ig}\), with RNA rows and \(i\in T_0\) columns. Center each condition column across RNA rows and obtain rank-\(r\) PCA scores \(Z=U_r\Sigma_r\). The descriptor of perturbation \(i\) is row \(Z_{j(i),:}\); the control descriptor is zero. Fit coordinate means and standard deviations on these original training-target descriptors plus that zero-control row, and apply the same standardization to every query descriptor. A constant descriptor coordinate has unit divisor.
|
| 220 |
+
|
| 221 |
+
For standardized descriptors \(x\), use a positive bandwidth \(\ell>0\) and positive ridge strength \(\lambda_R>0\), and define \(k_R(x,x')=\exp(-\|x-x'\|^2/\ell^2)\). Let \(K\) be its matrix on \(T_0\), \(H=I-\mathbf1\mathbf1^\top/|T_0|\), and \(K_c=HKH\). The centered query row is
|
| 222 |
+
|
| 223 |
+
\[
|
| 224 |
+
k_c(q,i)=k_R(q,i)-|T_0|^{-1}\sum_{a\in T_0}k_R(q,a)
|
| 225 |
+
-|T_0|^{-1}\sum_{a\in T_0}K_{ai}
|
| 226 |
+
+|T_0|^{-2}\sum_{a,d\in T_0}K_{ad}.
|
| 227 |
+
\]
|
| 228 |
+
|
| 229 |
+
With \(\mu^R=|T_0|^{-1}\sum_{i\in T_0}Y_i\), this view is
|
| 230 |
+
|
| 231 |
+
\[
|
| 232 |
+
B^R_q=\mu^R+k_c(q,:)(K_c+\lambda_R I)^{-1}(Y_{T_0}-\mathbf1(\mu^R)^\top).
|
| 233 |
+
\]
|
| 234 |
+
|
| 235 |
+
The RBF response origin \(\mu^R\) includes control and differs from \(b\). For any fixed numerical settings, this entire view remains fitted on \(T_0\) when validation rows enter another view's atlas.
|
| 236 |
+
|
| 237 |
+
**Ordinary response mean.** For the applicable noncontrol atlas pool,
|
| 238 |
+
|
| 239 |
+
\[
|
| 240 |
+
B^M_{qg}=|A|^{-1}\sum_{i\in A}Y_{ig}.
|
| 241 |
+
\]
|
| 242 |
+
|
| 243 |
+
**Downstream-weighted global mean.** For a positive tempering exponent \(\alpha\),
|
| 244 |
+
|
| 245 |
+
\[
|
| 246 |
+
B^W_{qg}=\frac{\sum_{i\in A}W_{ig}^{\alpha}Y_{ig}}
|
| 247 |
+
{\sum_{i\in A}W_{ig}^{\alpha}}.
|
| 248 |
+
\]
|
| 249 |
+
|
| 250 |
+
A zero denominator falls back to \(b_g\). Weights are indexed by reference condition and output RNA. They are not a query-specific vector of scalar reference weights broadcast over all outputs.
|
| 251 |
+
|
| 252 |
+
**Symbol-family weighted mean.** Let \(f(s)\) be the complete case-preserving initial run of ASCII letters in symbol \(s\), or the entire symbol if there is no initial letter run. Set \(F_q=\{i\in A:f(s_{j(i)})=f(s_{j(q)})\}\). Then
|
| 253 |
+
|
| 254 |
+
\[
|
| 255 |
+
B^F_{qg}=\frac{\sum_{i\in F_q}W_{ig}^{\alpha}Y_{ig}}
|
| 256 |
+
{\sum_{i\in F_q}W_{ig}^{\alpha}}.
|
| 257 |
+
\]
|
| 258 |
+
|
| 259 |
+
An absent family or zero denominator falls back to \(B^W_{qg}\). The same exponent \(\alpha\) is used for global and family means; keys are compared in full, without prefix truncation.
|
| 260 |
+
|
| 261 |
+
**Character inner-product kernel ridge regression.** Fit a case-sensitive character n-gram vocabulary and inverse-document frequencies on original training-target symbols. Use sublinear term frequency and unit-Euclidean-norm TF-IDF vectors \(\phi_i\) for nonzero feature rows; zero feature rows remain zero. Transform query symbols with the same fitted vocabulary. Use a positive ridge strength \(\lambda_C>0\). With \(K^C_{ia}=\phi_i^\top\phi_a\) for \(i,a\in A\) and \(k^C(q,i)=\phi_q^\top\phi_i\),
|
| 262 |
+
|
| 263 |
+
\[
|
| 264 |
+
B^C_q=b+k^C(q,:)(K^C+\lambda_C I)^{-1}
|
| 265 |
+
(\widetilde Y_A-\mathbf1b^\top).
|
| 266 |
+
\]
|
| 267 |
+
|
| 268 |
+
This is a ridge regression operator, not a similarity-normalized mean. For calibration, use the kernel on the atlas with the calibration condition removed. An exact kernel leave-one-out identity or an equivalent refit implements that same exclusion.
|
| 269 |
+
|
| 270 |
+
**Incoming-profile weighted neighbors.** Write \(t_{ig}\) for the signed raw Welch statistic of readout \(g\) under original training intervention \(i\), before the absolute value, min-max scaling, squaring and own-target masking used to construct \(W\). For a training target readout \(g=j(i)\), replace its own-intervention entry by the mean over its other original training interventions. Apply \(u_{gi}=\operatorname{sign}(t_{ig})\log(1+|t_{ig}|)\), center each readout's profile across intervention coordinates, and normalize it to unit Euclidean length. A constant profile becomes zero. Denote the resulting fixed descriptor by \(p_g\).
|
| 271 |
+
|
| 272 |
+
Choose the \(k\) most similar reference targets in \(A\) under \(s_{qi}=p_{j(q)}^\top p_{j(i)}\), calling the set \(N_q\). For positive temperature \(\tau\) and a positive exponent \(\beta\),
|
| 273 |
+
|
| 274 |
+
\[
|
| 275 |
+
a_{qi}=\exp\{(s_{qi}-\max_{d\in N_q}s_{qd})/\tau\},\qquad
|
| 276 |
+
B^N_{qg}=\frac{\sum_{i\in N_q}a_{qi}W_{ig}^{\beta}\widetilde Y_{ig}}
|
| 277 |
+
{\sum_{i\in N_q}a_{qi}W_{ig}^{\beta}}.
|
| 278 |
+
\]
|
| 279 |
+
|
| 280 |
+
A zero denominator falls back to \(B^W_{qg}\). Exclude a calibration condition before neighbor selection. The neighbor exponent \(\beta\) is distinct from the shared global/family exponent \(\alpha\).
|
| 281 |
+
|
| 282 |
+
**Symbol-hierarchy kernel ridge with an unpenalized intercept.** Choose a positive prefix length \(p\). Map each case-preserved target symbol to namespace-separated binary features for: its first \(p\) characters; its alphabetic characters concatenated in their original order; its complete initial ASCII-uppercase-letter run; its first run from the sequence of maximal ASCII-uppercase-letter runs and maximal digit runs, taken separately in encounter order; and the concatenation of the first two runs in that sequence. Alphabetic-character extraction preserves Unicode alphabetic characters; the run extraction separates letter runs from digit runs, equivalent to `[A-Z]+|[0-9]+` rather than `[A-Z0-9]+`. Include the last feature only when two runs exist, and the first-run feature only when a run exists. These feature groups have distinct namespaces even when their string values coincide. Each present dictionary key has value one, rather than a frequency count. The supplied symbols used for this map must have an initial uppercase run.
|
| 283 |
+
|
| 284 |
+
Normalize each nonzero feature row to unit Euclidean length, writing it as \(\psi_i\). The feature dictionary may cover every symbol in the supplied annotations: unused coordinates are zero and no response labels determine it. With a positive ridge strength \(\lambda_H\), define \(K^H_{ia}=\psi_i^\top\psi_a\) on the applicable atlas \(A\). Let \(c^0_g\) denote the original training-control mean and define a distinct response copy
|
| 285 |
+
|
| 286 |
+
\[
|
| 287 |
+
Y^H_{ig}=\begin{cases}c^0_g,&g=j(i),\\Y_{ig},&g\ne j(i).\end{cases}
|
| 288 |
+
\]
|
| 289 |
+
|
| 290 |
+
This replaces an own-target entry with the control mean, not with \(b_g\). Fit kernel ridge with an unpenalized intercept. One equivalent expression is
|
| 291 |
+
|
| 292 |
+
\[
|
| 293 |
+
Q=(K^H+\lambda_H I)^{-1},\quad u=Q\mathbf1,\quad d=\mathbf1^\top u,\quad
|
| 294 |
+
M=Q-uu^\top/d,\quad C_H=MY^H_A,\quad b_H=(\mathbf1^\top QY^H_A)/d,
|
| 295 |
+
\]
|
| 296 |
+
|
| 297 |
+
\[
|
| 298 |
+
B^H_q=b_H+k^H(q,:)C_H,\qquad k^H(q,i)=\psi_q^\top\psi_i.
|
| 299 |
+
\]
|
| 300 |
+
|
| 301 |
+
For calibration, remove the condition before fitting or use the equivalent intercept-corrected leave-one-out identity on the full legal atlas:
|
| 302 |
+
|
| 303 |
+
\[
|
| 304 |
+
B^{H,\mathrm{LOO}}_i=Y^H_i-(C_H)_i/M_{ii}.
|
| 305 |
+
\]
|
| 306 |
+
|
| 307 |
+
Apply this identity to the raw response prediction before nonnegative clipping. Require positive finite denominators; a zero query kernel follows the fitted intercept, rather than an unrelated mean fallback. This operator differs from the preceding fixed-origin TF-IDF ridge in both its annotation features and its unpenalized intercept.
|
| 308 |
+
|
| 309 |
+
**Projected local character-count neighbors.** Form case-sensitive character n-gram count vectors over a selected positive n-gram-length range, counting repeated substrings. Use ordinary term counts with neither inverse-document-frequency weighting nor sublinear term-frequency transformation, then normalize nonzero vectors to unit Euclidean length; zero rows remain zero. The annotation-only vocabulary may include every supplied gene symbol. Let these vectors be \(\chi_i\) and let \(r_{qi}=\chi_q^\top\chi_i\), a nonnegative cosine similarity.
|
| 310 |
+
|
| 311 |
+
Choose up to \(k_L\) nearest reference conditions in the applicable atlas, excluding a calibration condition before selection. Fix a deterministic tie rule. With positive power \(\eta\), define
|
| 312 |
+
|
| 313 |
+
\[
|
| 314 |
+
\omega_{qi}=\frac{r_{qi}^{\eta}}{\sum_{a\in N^L_q}r_{qa}^{\eta}},\qquad
|
| 315 |
+
\overline Y^L_q=\sum_{i\in N^L_q}\omega_{qi}Y_i.
|
| 316 |
+
\]
|
| 317 |
+
|
| 318 |
+
These reference responses are measured \(Y\), with no own-target replacement and no downstream-weight multiplier. Zero similarity-weight mass falls back to \(b\). From the original noncontrol training matrix \(Y_T-\mathbf1b^\top\), compute orthonormal right singular vectors and retain a chosen rank \(r_L\) within its numerical rank, collected as rows of \(V_L\). Do not standardize output columns, whiten the basis or multiply it by singular values. The view is
|
| 319 |
+
|
| 320 |
+
\[
|
| 321 |
+
B^L_q=b+(\overline Y^L_q-b)V_L^\top V_L.
|
| 322 |
+
\]
|
| 323 |
+
|
| 324 |
+
A zero retained rank returns \(b\). This output projection is fitted only on original training responses and remains fixed when validation rows enter or leave the atlas. It differs from incoming-profile neighbors: neighbor identity comes from character counts, while the training-only projection regularizes the transferred output response.
|
| 325 |
+
|
| 326 |
+
Project each of the eight base expression predictions onto nonnegative values before calibration. Numerical zero-mass and zero-norm tolerances remain implementation choices; preserve the stated fallbacks. Below, \(B^{(a)}\) denotes these projected base predictions, with one fixed ordering of the eight components.
|
| 327 |
+
|
| 328 |
+
### Three sequential calibration fits
|
| 329 |
+
|
| 330 |
+
Build leave-one-condition calibration predictions for every \(c\in C\) using its applicable atlas exclusion. Set the eight-component vector and centered target to
|
| 331 |
+
|
| 332 |
+
\[
|
| 333 |
+
X_{cg}=(B^{(a)}_{cg}-b_g)_{a=1}^{8},\qquad z_{cg}=Y_{cg}-b_g.
|
| 334 |
+
\]
|
| 335 |
+
|
| 336 |
+
All calibration observations and their weights come from \(C\). Define
|
| 337 |
+
|
| 338 |
+
\[
|
| 339 |
+
G_g=|C|^{-1}\sum_{c\in C}W_{cg}X_{cg}X_{cg}^\top,\quad
|
| 340 |
+
h_g=|C|^{-1}\sum_{c\in C}W_{cg}X_{cg}z_{cg},\quad
|
| 341 |
+
G=\sum_g G_g,\quad h=\sum_g h_g.
|
| 342 |
+
\]
|
| 343 |
+
|
| 344 |
+
Let \(s=\max(\operatorname{tr}(G)/8,\epsilon)\), \(s_g=\max(\operatorname{tr}(G_g)/8,\epsilon)\), with a positive numerical floor \(\epsilon\), and let \(v_0=\mathbf1/8\). Fit, in this order,
|
| 345 |
+
|
| 346 |
+
\[
|
| 347 |
+
v_S=\arg\min_{v\ge0,\;\mathbf1^\top v=1}
|
| 348 |
+
\{v^\top Gv-2h^\top v+\lambda_Ss\|v-v_0\|^2\},
|
| 349 |
+
\]
|
| 350 |
+
|
| 351 |
+
\[
|
| 352 |
+
v_A=\arg\min_{v\in\mathbb R^8}
|
| 353 |
+
\{v^\top Gv-2h^\top v+\lambda_As\|v-v_S\|^2\},
|
| 354 |
+
\]
|
| 355 |
+
|
| 356 |
+
\[
|
| 357 |
+
v_g=\arg\min_{v\in\mathbb R^8}
|
| 358 |
+
\{v^\top G_gv-2h_g^\top v+\lambda_Gs_g\|v-v_A\|^2\}.
|
| 359 |
+
\]
|
| 360 |
+
|
| 361 |
+
The Gram forms express the corresponding condition/output-weighted squared losses, up to constants independent of the coefficients. Only \(v_S\) lies on a simplex. The global signed vector \(v_A\) and output-specific signed vectors \(v_g\) have neither a nonnegativity nor a sum-to-one constraint. The positive strengths \(\lambda_S,\lambda_A,\lambda_G\) are numerical choices; \(\lambda_G\) is a shared scalar with output-specific scales \(s_g\). These are sequential fits toward already fitted priors, not one joint optimization over all stages. The fixed origin supplies the response offset; there is no additional fitted intercept in these equations.
|
| 362 |
+
|
| 363 |
+
For a final query, the delivered expression is
|
| 364 |
+
|
| 365 |
+
\[
|
| 366 |
+
\widehat Y_{qg}=\max\{0,\ b_g+v_g^\top(B^{(a)}_{qg}-b_g)_{a=1}^{8}\}.
|
| 367 |
+
\]
|
| 368 |
+
|
| 369 |
+
This equation defines the final combination of the eight views. Implement the method with your selected numerical parameters and return its predictions in the common required order.
|
| 370 |
+
|
| 371 |
+
### Public selection and validation boundaries
|
| 372 |
+
|
| 373 |
+
Choose numerical settings and a declared finite comparison using public validation only. In particular, select the global-signed and output-signed shrinkage strengths jointly. Declare the candidate ordering and tie rule before comparison. Fold counts, membership rules, seeds, candidate values and numerical solvers are not supplied here.
|
| 374 |
+
|
| 375 |
+
For a validation split with fitting pool \(C\) and held-out pool \(\mathcal H\), build \(A_C\), its calibration leave-one-out bank and all three coefficient fits using \(C\) alone. Predict \(\mathcal H\) as queries outside that atlas. Its response and weight rows enter only the evaluation of these candidate predictions, not the atlas or coefficient fits. Inner held-out evaluations can select settings within the outer-training pool; outer held-out evaluations remain outside that selection. For any fixed numerical settings, original-training descriptors, the original-training RBF and the original-training output projection remain fixed throughout; symbol-only feature dictionaries do not use validation responses.
|
| 376 |
+
|
| 377 |
+
For an outer assessment, repeat the entire parameter-selection procedure within the outer-training validation pool. Every inner fit excludes both the outer held-out rows and its inner held-out rows. Score the selected procedure once on the outer held-out fold with its supplied condition/output weights. Store and align condition identifiers, output identifiers and fold membership with prediction banks. Use the same component/calibration functions for held-out predictions and final queries, changing only the permitted response pool. A bank built using all validation labels does not become outer-clean merely by deleting the held-out rows afterward.
|
| 378 |
+
|
| 379 |
+
If a residual correction is independently investigated outside this specified estimator, it requires residuals rebuilt within each outer-training pool and evaluation together with that correction on the outer-held-out rows. It is not part of the final eight-view equation above. After public selection, refit the selected eight-view procedure on all permitted training and validation observations and execute it for the query identifiers. Repeated selection on outer results makes those results development evidence; fitting loss and deterministic replay alone do not establish independent predictive quality.
|
tasks/biology.unseen_perturbation_rna_response__seed31415/prompt_b3.md
ADDED
|
@@ -0,0 +1,192 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Predict downstream RNA responses to unseen gene perturbations
|
| 2 |
+
|
| 3 |
+
Predict the mean measured RNA response for 233 held-out single-gene CRISPR interference conditions in K562 cells. Training observations cover other perturbed genes. The scientific target is the downstream expression response: the directly targeted gene remains in the output, but its weight for its own condition is zero. Return your best complete finite prediction array, including when full credit is not reached. This is prediction from existing measurements, not experimental intervention design.
|
| 4 |
+
|
| 5 |
+
## Work allowance
|
| 6 |
+
|
| 7 |
+
Read `timeout_seconds` in `task_info.json`, record your start time and manage the remaining development allowance. Reasoning, exploration, implementation and verification use that allowance. It is separate from fresh-program replay below. Produce a runnable program and complete output early, retain a working result while improving it, and finish normally within the allowance.
|
| 8 |
+
|
| 9 |
+
## Measurements and fixed split
|
| 10 |
+
|
| 11 |
+
The source is the Replogle K562 essential-gene CRISPRi experiment, sampled six days after perturbation. Every retained matrix feature was matched by its Ensembl ID to original Gene Expression annotation. The released count matrix contains 8,563 verified endogenous RNA features; upstream authors had already restricted their release to genes averaging more than 0.01 UMI per cell. Guide-capture features and other modalities are not expression outputs.
|
| 12 |
+
|
| 13 |
+
For each cell and released RNA gene, the value is
|
| 14 |
+
|
| 15 |
+
`z = ln(1 + 10000 * raw_gene_count / sum_of_all_8563_released_RNA_counts_in_that_cell)`.
|
| 16 |
+
|
| 17 |
+
Normalization precedes output-panel selection. The denominator includes all 8,563 released RNA features, not only the 3,323 selected output features. For each test condition, the target is the gene-wise mean of z over its fixed measured GT-half cells. These are experimental statistics, not predictions produced by an author model.
|
| 18 |
+
|
| 19 |
+
The cohort contains 1,546 canonical target Ensembl IDs with a unique measured RNA identity and at least 64 cells each. The eligible IDs were assigned once to fixed, gene-disjoint training, validation and test sets of 1,082 / 231 / 233 targets. All cells and guides belonging to a target stay in the same split. Training has 188,047 cells including 7,483 controls. Validation and test each have 1,604 control cells.
|
| 20 |
+
|
| 21 |
+
Within each condition, cells are split into two fixed random halves. The first half supplies GT statistics; the other supplies a private technical comparison. The 48 GEM groups and cell halves are technical groupings, not independent biological replicates. The 3,323-feature panel combines 2,000 highly variable RNA genes selected using training cells only with all 1,546 eligible target RNA genes. Every validation and test target therefore occurs as an observed readout column in training data; its held-out perturbation response does not occur there. Target identity is `obs.gene_id` matched to `var.ensembl_id`, rather than a guessed symbol alias.
|
| 22 |
+
|
| 23 |
+
## Inputs
|
| 24 |
+
|
| 25 |
+
All public inputs are in `data/`. Load NPZ files with `allow_pickle=False`. They contain only numeric or Unicode arrays. Preserve all declared row and column orders. Each NPZ array below states its dtype and shape; every JSON field type is listed in the field-schema sections at the end.
|
| 26 |
+
|
| 27 |
+
- `README.md`: text description of the public files and scientific contract.
|
| 28 |
+
- `train_condition_stats.npz`: `condition_ids` is Unicode[1083], with `control` first and 1,082 training target Ensembl IDs thereafter; `mean_expression` and `variance` are float64[1083,3323]; `n_cells` is int64[1083]. Variance is the per-cell sample variance, with ddof=1, of the normalized log expression. Means and variances use only training cells.
|
| 29 |
+
- `validation_targets.npz`: `condition_ids` is Unicode[231]; `mean_expression` and `metric_weights` are float64[231,3323]; `n_GT` is int64[231]. These measured validation GT-half means and downstream scoring weights are available for tuning and model selection. Technical second-half values are not supplied.
|
| 30 |
+
- `test_queries.json`: `condition_ids` is the ordered list of 233 test target Ensembl strings; `shape` is `[233,3323]`. This file contains identifiers and the requested shape only.
|
| 31 |
+
- `genes.json`: `gene_ids` and `gene_symbols` are ordered lists of 3,323 strings; `raw_column_indices` is the corresponding list of integer indices in the released RNA matrix; `target_id_semantics` is a descriptive string. `gene_ids` defines the output-column order. Symbols are annotations; Ensembl IDs define identity.
|
| 32 |
+
- `metadata.json`: a JSON object defining the dataset, expression scale, split semantics, feature panel, file list, scoring constants and formula, output schema, and resource/data-access contract. Shapes are integer arrays; `scientific_scoring`, `output` and `resource_contract` are explanatory objects.
|
| 33 |
+
- `provenance.json`: dataset citation, license, upstream filtering, feature-ID audit, split summary and public-file hashes. It adds no expression observations. Original full-data gene summary statistics are not supplied as predictive inputs.
|
| 34 |
+
- `manifest.json`: `schema_version` is an integer; `files` maps other public filenames to `sha256` strings and byte counts; `public_test_content` describes the identifier-only test data; `private_evaluation_arrays_included` is false.
|
| 35 |
+
|
| 36 |
+
## Downstream weights and scientific score
|
| 37 |
+
|
| 38 |
+
For each validation or test condition, construct weights from that condition's GT-half cells and the corresponding split's control GT-half cells. Let m, s² and n denote a gene's mean, sample variance and cell count. Compute
|
| 39 |
+
|
| 40 |
+
`t_g = (m_condition,g - m_control,g) / sqrt(s²_condition,g/n_condition + s²_control,g/n_control)`.
|
| 41 |
+
|
| 42 |
+
Set t_g=0 when the denominator is zero. Across all 3,323 genes, min-max scale abs(t_g) to [0,1] and square it. Then set the condition's own canonical target Ensembl gene's entry to exactly zero and renormalize the remaining entries to sum to one. Every frozen condition has positive remaining weight. A degenerate author weight row is an evaluator data error, not a submitted prediction failure. Technical second-half cells never determine these weights.
|
| 43 |
+
|
| 44 |
+
For prediction P, private measured GT means Y and private weights w, the loss is
|
| 45 |
+
|
| 46 |
+
`L = (1/233) * sum_c sum_g w[c,g] * (P[c,g] - Y[c,g])²`.
|
| 47 |
+
|
| 48 |
+
Sum weighted squared errors over genes; do not take an additional gene mean. The score is
|
| 49 |
+
|
| 50 |
+
`score = clip(100 * (B - L) / (B - T), 0, 100)`,
|
| 51 |
+
|
| 52 |
+
`B = 0.03441638519456471`,
|
| 53 |
+
|
| 54 |
+
`T = 0.00714613515111071`.
|
| 55 |
+
|
| 56 |
+
B is the test loss of the equal-weight mean response of the 1,082 noncontrol training conditions. T is the loss of the private technical second-half means against the measured GT-half means, using the same downstream weights. T calibrates sampling reproducibility in this cohort; it is not an independent biological replication result or a proven prediction limit. Losses better than T also receive 100. The evaluator retains raw loss and diagnostics. No method is required to reproduce private experimental labels exactly. Expected range: 100 marks technical half-sample reproducibility, not an attainable target. The strongest method known to the task author scores about 38; scores above about 40 are not expected.
|
| 57 |
+
|
| 58 |
+
## Output
|
| 59 |
+
|
| 60 |
+
Deliver a self-contained `analysis.py` of at most 65,536 bytes (at most 12,288 bytes after zlib level-9 compression, at most 8,192 characters in total across all string, bytes, f-string, numeric and comment tokens, and at most 1,024 characters in any single such token) and `results/predictions.npy` of at most 8,388,608 bytes. The NPY array must have dtype float32 or float64, shape `[233,3323]`, and finite values. Row order is `test_queries.json.condition_ids`; column order is `genes.json.gene_ids`. Include the own-target column even though its condition-specific scoring weight is zero. Finite values of either sign are accepted; there is no extra sign, norm or accuracy gate.
|
| 61 |
+
|
| 62 |
+
`python analysis.py` must recompute your prediction array using `data/` in a fresh workspace. Embed all helper code in that one source file. A source exceeding any size, compression or literal limit above fails the hard gate without being replayed; compute values from `data/` rather than embedding them. Do not depend on a development checkpoint, cached predictions, private labels, or previous scratch files. The hard replay-consistency gate checks both arrays' shape, dtype and finiteness and requires every element to satisfy
|
| 63 |
+
|
| 64 |
+
`abs(submitted - regenerated) <= 1e-5 + 1e-6 * abs(regenerated)`.
|
| 65 |
+
|
| 66 |
+
This is `np.allclose(submitted, regenerated, atol=1e-5, rtol=1e-6)`. It compares your submitted array to your fresh computation, not to experimental truth. Your submission must pass this gate separately from the scientific score.
|
| 67 |
+
|
| 68 |
+
Use only the supplied public data for scientific learning. Training and validation responses may be used with any available computational method or installed library. Do not fetch external datasets, pretrained models, online test responses, or external knowledge resources. Connectivity required by the agent provider is separate from the scientific-data boundary. The same rule applies at all prompt levels.
|
| 69 |
+
|
| 70 |
+
Fresh replay has 1,200 seconds, four assigned CPUs, four numerical-library threads and 8 GiB memory with no swap. Preserve assigned CPU affinity and run offline. The runtime provides Python>=3.11,<3.12, NumPy1.24.3, SciPy1.11.4, pandas1.5.3 and scikit-learn1.3.2 with dependencies. Temporary work belongs in the supplied writable scratch location. Inputs are immutable. The evaluator pre-creates writable `results/predictions.npy`; overwrite it directly using `np.save` on a binary file handle. Do not rename it or create additional result files.
|
| 71 |
+
|
| 72 |
+
A completed low-scoring prediction remains a valid measured attempt. An external agent timeout or evaluator infrastructure failure is not a valid scientific difficulty result. Finish with both the program and completed prediction file.
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
## Complete JSON metadata fields
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
## Additional public JSON field schemas
|
| 79 |
+
|
| 80 |
+
The following lists every field and its type in these three public JSON files. `array[n]` specifies the array length.
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
`data/metadata.json` (object):
|
| 84 |
+
|
| 85 |
+
```text
|
| 86 |
+
"dataset": string
|
| 87 |
+
"expression": string
|
| 88 |
+
"target_semantics": string
|
| 89 |
+
"train_stats": string
|
| 90 |
+
"genes": string
|
| 91 |
+
"validation": string
|
| 92 |
+
"test": string
|
| 93 |
+
"primary_metric": string
|
| 94 |
+
"no_test_response_in_gene_descriptors": boolean
|
| 95 |
+
"files": array[8] of string
|
| 96 |
+
"output_shape": array[2] of integer
|
| 97 |
+
"task_id": string
|
| 98 |
+
"schema_version": integer
|
| 99 |
+
"cell_line": string
|
| 100 |
+
"perturbation_type": string
|
| 101 |
+
"normalization_denominator_features": integer
|
| 102 |
+
"feature_selection": string
|
| 103 |
+
"scientific_scoring": object
|
| 104 |
+
"primary_loss": string
|
| 105 |
+
"own_target_weight": string
|
| 106 |
+
"weight_construction": string
|
| 107 |
+
"score_formula": string
|
| 108 |
+
"B": number
|
| 109 |
+
"T": number
|
| 110 |
+
"B_definition": string
|
| 111 |
+
"T_definition": string
|
| 112 |
+
"GT_origin": string
|
| 113 |
+
"technical_half_limit": string
|
| 114 |
+
"output": object
|
| 115 |
+
"path": string
|
| 116 |
+
"dtype": array[2] of string
|
| 117 |
+
"shape": array[2] of integer
|
| 118 |
+
"finite": boolean
|
| 119 |
+
"maximum_bytes": integer
|
| 120 |
+
"source_path": string
|
| 121 |
+
"maximum_source_bytes": integer
|
| 122 |
+
"maximum_compressed_source_bytes": integer
|
| 123 |
+
"maximum_source_literal_characters": integer
|
| 124 |
+
"maximum_single_literal_characters": integer
|
| 125 |
+
"literal_definition": string
|
| 126 |
+
"resource_contract": object
|
| 127 |
+
"fresh_replay_seconds": integer
|
| 128 |
+
"CPUs": integer
|
| 129 |
+
"threads": integer
|
| 130 |
+
"memory_GiB": integer
|
| 131 |
+
"development_budget": string
|
| 132 |
+
"scientific_data_access": string
|
| 133 |
+
"replay_atol": number
|
| 134 |
+
"replay_rtol": number
|
| 135 |
+
```
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
`data/provenance.json` (object):
|
| 139 |
+
|
| 140 |
+
```text
|
| 141 |
+
"schema_version": integer
|
| 142 |
+
"dataset": string
|
| 143 |
+
"license": string
|
| 144 |
+
"original_paper": string
|
| 145 |
+
"feature_identity_audit": string
|
| 146 |
+
"source_prefilter": string
|
| 147 |
+
"target_identity": string
|
| 148 |
+
"eligible_cohort": string
|
| 149 |
+
"split": string
|
| 150 |
+
"technical_halves": string
|
| 151 |
+
"training_cells": integer
|
| 152 |
+
"training_controls": integer
|
| 153 |
+
"validation_controls": integer
|
| 154 |
+
"test_controls": integer
|
| 155 |
+
"public_numeric_source_hashes": object
|
| 156 |
+
"train_condition_stats.npz": string
|
| 157 |
+
"validation_targets.npz": string
|
| 158 |
+
"offline_boundary": string
|
| 159 |
+
"no_original_full_data_summary_priors": string
|
| 160 |
+
```
|
| 161 |
+
|
| 162 |
+
|
| 163 |
+
`data/manifest.json` (object):
|
| 164 |
+
|
| 165 |
+
```text
|
| 166 |
+
"schema_version": integer
|
| 167 |
+
"files": object
|
| 168 |
+
"README.md": object
|
| 169 |
+
"sha256": string
|
| 170 |
+
"bytes": integer
|
| 171 |
+
"genes.json": object
|
| 172 |
+
"sha256": string
|
| 173 |
+
"bytes": integer
|
| 174 |
+
"metadata.json": object
|
| 175 |
+
"sha256": string
|
| 176 |
+
"bytes": integer
|
| 177 |
+
"provenance.json": object
|
| 178 |
+
"sha256": string
|
| 179 |
+
"bytes": integer
|
| 180 |
+
"test_queries.json": object
|
| 181 |
+
"sha256": string
|
| 182 |
+
"bytes": integer
|
| 183 |
+
"train_condition_stats.npz": object
|
| 184 |
+
"sha256": string
|
| 185 |
+
"bytes": integer
|
| 186 |
+
"validation_targets.npz": object
|
| 187 |
+
"sha256": string
|
| 188 |
+
"bytes": integer
|
| 189 |
+
"public_test_content": string
|
| 190 |
+
"private_evaluation_arrays_included": boolean
|
| 191 |
+
```
|
| 192 |
+
|
tasks/biology.unseen_perturbation_rna_response__seed31415/prompt_b4.md
ADDED
|
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Predict downstream RNA responses to unseen gene perturbations
|
| 2 |
+
|
| 3 |
+
Predict the mean measured RNA response for 233 held-out single-gene CRISPR interference conditions in K562 cells. Training observations cover other perturbed genes. The scientific target is the downstream expression response: the directly targeted gene remains in the output, but its weight for its own condition is zero. Return your best complete finite prediction array, including when full credit is not reached. This is prediction from existing measurements, not experimental intervention design.
|
| 4 |
+
|
| 5 |
+
## Background
|
| 6 |
+
|
| 7 |
+
The two NPZ inputs are DEFLATE-compressed ZIP archives of NumPy `.npy` arrays, and every JSON input is plain ASCII text. Each `.npy` member stores a short header with dtype, shape and memory order, followed by the raw array bytes.
|
| 8 |
+
|
| 9 |
+
## Work allowance
|
| 10 |
+
|
| 11 |
+
Read `timeout_seconds` in `task_info.json`, record your start time and manage the remaining development allowance. Reasoning, exploration, implementation and verification use that allowance. It is separate from fresh-program replay below. Produce a runnable program and complete output early, retain a working result while improving it, and finish normally within the allowance.
|
| 12 |
+
|
| 13 |
+
## Measurements and fixed split
|
| 14 |
+
|
| 15 |
+
The source is the Replogle K562 essential-gene CRISPRi experiment, sampled six days after perturbation. Every retained matrix feature was matched by its Ensembl ID to original Gene Expression annotation. The released count matrix contains 8,563 verified endogenous RNA features; upstream authors had already restricted their release to genes averaging more than 0.01 UMI per cell. Guide-capture features and other modalities are not expression outputs.
|
| 16 |
+
|
| 17 |
+
For each cell and released RNA gene, the value is
|
| 18 |
+
|
| 19 |
+
`z = ln(1 + 10000 * raw_gene_count / sum_of_all_8563_released_RNA_counts_in_that_cell)`.
|
| 20 |
+
|
| 21 |
+
Normalization precedes output-panel selection. The denominator includes all 8,563 released RNA features, not only the 3,323 selected output features. For each test condition, the target is the gene-wise mean of z over its fixed measured GT-half cells. These are experimental statistics, not predictions produced by an author model.
|
| 22 |
+
|
| 23 |
+
The cohort contains 1,546 canonical target Ensembl IDs with a unique measured RNA identity and at least 64 cells each. The eligible IDs were assigned once to fixed, gene-disjoint training, validation and test sets of 1,082 / 231 / 233 targets. All cells and guides belonging to a target stay in the same split. Training has 188,047 cells including 7,483 controls. Validation and test each have 1,604 control cells.
|
| 24 |
+
|
| 25 |
+
Within each condition, cells are split into two fixed random halves. The first half supplies GT statistics; the other supplies a private technical comparison. The 48 GEM groups and cell halves are technical groupings, not independent biological replicates. The 3,323-feature panel combines 2,000 highly variable RNA genes selected using training cells only with all 1,546 eligible target RNA genes. Every validation and test target therefore occurs as an observed readout column in training data; its held-out perturbation response does not occur there. Target identity is `obs.gene_id` matched to `var.ensembl_id`, rather than a guessed symbol alias.
|
| 26 |
+
|
| 27 |
+
## Inputs
|
| 28 |
+
|
| 29 |
+
All public inputs are in `data/`. Load NPZ files with `allow_pickle=False`. They contain only numeric or Unicode arrays. Preserve all declared row and column orders. Each NPZ array below states its dtype and shape; every JSON field type is listed in the field-schema sections at the end.
|
| 30 |
+
|
| 31 |
+
- `README.md`: text description of the public files and scientific contract.
|
| 32 |
+
- `train_condition_stats.npz`: `condition_ids` is Unicode[1083], with `control` first and 1,082 training target Ensembl IDs thereafter; `mean_expression` and `variance` are float64[1083,3323]; `n_cells` is int64[1083]. Variance is the per-cell sample variance, with ddof=1, of the normalized log expression. Means and variances use only training cells.
|
| 33 |
+
- `validation_targets.npz`: `condition_ids` is Unicode[231]; `mean_expression` and `metric_weights` are float64[231,3323]; `n_GT` is int64[231]. These measured validation GT-half means and downstream scoring weights are available for tuning and model selection. Technical second-half values are not supplied.
|
| 34 |
+
- `test_queries.json`: `condition_ids` is the ordered list of 233 test target Ensembl strings; `shape` is `[233,3323]`. This file contains identifiers and the requested shape only.
|
| 35 |
+
- `genes.json`: `gene_ids` and `gene_symbols` are ordered lists of 3,323 strings; `raw_column_indices` is the corresponding list of integer indices in the released RNA matrix; `target_id_semantics` is a descriptive string. `gene_ids` defines the output-column order. Symbols are annotations; Ensembl IDs define identity.
|
| 36 |
+
- `metadata.json`: a JSON object defining the dataset, expression scale, split semantics, feature panel, file list, scoring constants and formula, output schema, and resource/data-access contract. Shapes are integer arrays; `scientific_scoring`, `output` and `resource_contract` are explanatory objects.
|
| 37 |
+
- `provenance.json`: dataset citation, license, upstream filtering, feature-ID audit, split summary and public-file hashes. It adds no expression observations. Original full-data gene summary statistics are not supplied as predictive inputs.
|
| 38 |
+
- `manifest.json`: `schema_version` is an integer; `files` maps other public filenames to `sha256` strings and byte counts; `public_test_content` describes the identifier-only test data; `private_evaluation_arrays_included` is false.
|
| 39 |
+
|
| 40 |
+
## Downstream weights and scientific score
|
| 41 |
+
|
| 42 |
+
For each validation or test condition, construct weights from that condition's GT-half cells and the corresponding split's control GT-half cells. Let m, s² and n denote a gene's mean, sample variance and cell count. Compute
|
| 43 |
+
|
| 44 |
+
`t_g = (m_condition,g - m_control,g) / sqrt(s²_condition,g/n_condition + s²_control,g/n_control)`.
|
| 45 |
+
|
| 46 |
+
Set t_g=0 when the denominator is zero. Across all 3,323 genes, min-max scale abs(t_g) to [0,1] and square it. Then set the condition's own canonical target Ensembl gene's entry to exactly zero and renormalize the remaining entries to sum to one. Every frozen condition has positive remaining weight. A degenerate author weight row is an evaluator data error, not a submitted prediction failure. Technical second-half cells never determine these weights.
|
| 47 |
+
|
| 48 |
+
For prediction P, private measured GT means Y and private weights w, the loss is
|
| 49 |
+
|
| 50 |
+
`L = (1/233) * sum_c sum_g w[c,g] * (P[c,g] - Y[c,g])²`.
|
| 51 |
+
|
| 52 |
+
Sum weighted squared errors over genes; do not take an additional gene mean. The score is
|
| 53 |
+
|
| 54 |
+
`score = clip(100 * (B - L) / (B - T), 0, 100)`,
|
| 55 |
+
|
| 56 |
+
`B = 0.03441638519456471`,
|
| 57 |
+
|
| 58 |
+
`T = 0.00714613515111071`.
|
| 59 |
+
|
| 60 |
+
B is the test loss of the equal-weight mean response of the 1,082 noncontrol training conditions. T is the loss of the private technical second-half means against the measured GT-half means, using the same downstream weights. T calibrates sampling reproducibility in this cohort; it is not an independent biological replication result or a proven prediction limit. Losses better than T also receive 100. The evaluator retains raw loss and diagnostics. No method is required to reproduce private experimental labels exactly. Expected range: 100 marks technical half-sample reproducibility, not an attainable target. The strongest method known to the task author scores about 38; scores above about 40 are not expected.
|
| 61 |
+
|
| 62 |
+
## Output
|
| 63 |
+
|
| 64 |
+
Deliver a self-contained `analysis.py` of at most 65,536 bytes (at most 12,288 bytes after zlib level-9 compression, at most 8,192 characters in total across all string, bytes, f-string, numeric and comment tokens, and at most 1,024 characters in any single such token) and `results/predictions.npy` of at most 8,388,608 bytes. The NPY array must have dtype float32 or float64, shape `[233,3323]`, and finite values. Row order is `test_queries.json.condition_ids`; column order is `genes.json.gene_ids`. Include the own-target column even though its condition-specific scoring weight is zero. Finite values of either sign are accepted; there is no extra sign, norm or accuracy gate.
|
| 65 |
+
|
| 66 |
+
`python analysis.py` must recompute your prediction array using `data/` in a fresh workspace. Embed all helper code in that one source file. A source exceeding any size, compression or literal limit above fails the hard gate without being replayed; compute values from `data/` rather than embedding them. Do not depend on a development checkpoint, cached predictions, private labels, or previous scratch files. The hard replay-consistency gate checks both arrays' shape, dtype and finiteness and requires every element to satisfy
|
| 67 |
+
|
| 68 |
+
`abs(submitted - regenerated) <= 1e-5 + 1e-6 * abs(regenerated)`.
|
| 69 |
+
|
| 70 |
+
This is `np.allclose(submitted, regenerated, atol=1e-5, rtol=1e-6)`. It compares your submitted array to your fresh computation, not to experimental truth. Your submission must pass this gate separately from the scientific score.
|
| 71 |
+
|
| 72 |
+
Use only the supplied public data for scientific learning. Training and validation responses may be used with any available computational method or installed library. Do not fetch external datasets, pretrained models, online test responses, or external knowledge resources. Connectivity required by the agent provider is separate from the scientific-data boundary. The same rule applies at all prompt levels.
|
| 73 |
+
|
| 74 |
+
Fresh replay has 1,200 seconds, four assigned CPUs, four numerical-library threads and 8 GiB memory with no swap. Preserve assigned CPU affinity and run offline. The runtime provides Python>=3.11,<3.12, NumPy1.24.3, SciPy1.11.4, pandas1.5.3 and scikit-learn1.3.2 with dependencies. Temporary work belongs in the supplied writable scratch location. Inputs are immutable. The evaluator pre-creates writable `results/predictions.npy`; overwrite it directly using `np.save` on a binary file handle. Do not rename it or create additional result files.
|
| 75 |
+
|
| 76 |
+
A completed low-scoring prediction remains a valid measured attempt. An external agent timeout or evaluator infrastructure failure is not a valid scientific difficulty result. Finish with both the program and completed prediction file.
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
## Complete JSON metadata fields
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
## Additional public JSON field schemas
|
| 83 |
+
|
| 84 |
+
The following lists every field and its type in these three public JSON files. `array[n]` specifies the array length.
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
`data/metadata.json` (object):
|
| 88 |
+
|
| 89 |
+
```text
|
| 90 |
+
"dataset": string
|
| 91 |
+
"expression": string
|
| 92 |
+
"target_semantics": string
|
| 93 |
+
"train_stats": string
|
| 94 |
+
"genes": string
|
| 95 |
+
"validation": string
|
| 96 |
+
"test": string
|
| 97 |
+
"primary_metric": string
|
| 98 |
+
"no_test_response_in_gene_descriptors": boolean
|
| 99 |
+
"files": array[8] of string
|
| 100 |
+
"output_shape": array[2] of integer
|
| 101 |
+
"task_id": string
|
| 102 |
+
"schema_version": integer
|
| 103 |
+
"cell_line": string
|
| 104 |
+
"perturbation_type": string
|
| 105 |
+
"normalization_denominator_features": integer
|
| 106 |
+
"feature_selection": string
|
| 107 |
+
"scientific_scoring": object
|
| 108 |
+
"primary_loss": string
|
| 109 |
+
"own_target_weight": string
|
| 110 |
+
"weight_construction": string
|
| 111 |
+
"score_formula": string
|
| 112 |
+
"B": number
|
| 113 |
+
"T": number
|
| 114 |
+
"B_definition": string
|
| 115 |
+
"T_definition": string
|
| 116 |
+
"GT_origin": string
|
| 117 |
+
"technical_half_limit": string
|
| 118 |
+
"output": object
|
| 119 |
+
"path": string
|
| 120 |
+
"dtype": array[2] of string
|
| 121 |
+
"shape": array[2] of integer
|
| 122 |
+
"finite": boolean
|
| 123 |
+
"maximum_bytes": integer
|
| 124 |
+
"source_path": string
|
| 125 |
+
"maximum_source_bytes": integer
|
| 126 |
+
"maximum_compressed_source_bytes": integer
|
| 127 |
+
"maximum_source_literal_characters": integer
|
| 128 |
+
"maximum_single_literal_characters": integer
|
| 129 |
+
"literal_definition": string
|
| 130 |
+
"resource_contract": object
|
| 131 |
+
"fresh_replay_seconds": integer
|
| 132 |
+
"CPUs": integer
|
| 133 |
+
"threads": integer
|
| 134 |
+
"memory_GiB": integer
|
| 135 |
+
"development_budget": string
|
| 136 |
+
"scientific_data_access": string
|
| 137 |
+
"replay_atol": number
|
| 138 |
+
"replay_rtol": number
|
| 139 |
+
```
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
`data/provenance.json` (object):
|
| 143 |
+
|
| 144 |
+
```text
|
| 145 |
+
"schema_version": integer
|
| 146 |
+
"dataset": string
|
| 147 |
+
"license": string
|
| 148 |
+
"original_paper": string
|
| 149 |
+
"feature_identity_audit": string
|
| 150 |
+
"source_prefilter": string
|
| 151 |
+
"target_identity": string
|
| 152 |
+
"eligible_cohort": string
|
| 153 |
+
"split": string
|
| 154 |
+
"technical_halves": string
|
| 155 |
+
"training_cells": integer
|
| 156 |
+
"training_controls": integer
|
| 157 |
+
"validation_controls": integer
|
| 158 |
+
"test_controls": integer
|
| 159 |
+
"public_numeric_source_hashes": object
|
| 160 |
+
"train_condition_stats.npz": string
|
| 161 |
+
"validation_targets.npz": string
|
| 162 |
+
"offline_boundary": string
|
| 163 |
+
"no_original_full_data_summary_priors": string
|
| 164 |
+
```
|
| 165 |
+
|
| 166 |
+
|
| 167 |
+
`data/manifest.json` (object):
|
| 168 |
+
|
| 169 |
+
```text
|
| 170 |
+
"schema_version": integer
|
| 171 |
+
"files": object
|
| 172 |
+
"README.md": object
|
| 173 |
+
"sha256": string
|
| 174 |
+
"bytes": integer
|
| 175 |
+
"genes.json": object
|
| 176 |
+
"sha256": string
|
| 177 |
+
"bytes": integer
|
| 178 |
+
"metadata.json": object
|
| 179 |
+
"sha256": string
|
| 180 |
+
"bytes": integer
|
| 181 |
+
"provenance.json": object
|
| 182 |
+
"sha256": string
|
| 183 |
+
"bytes": integer
|
| 184 |
+
"test_queries.json": object
|
| 185 |
+
"sha256": string
|
| 186 |
+
"bytes": integer
|
| 187 |
+
"train_condition_stats.npz": object
|
| 188 |
+
"sha256": string
|
| 189 |
+
"bytes": integer
|
| 190 |
+
"validation_targets.npz": object
|
| 191 |
+
"sha256": string
|
| 192 |
+
"bytes": integer
|
| 193 |
+
"public_test_content": string
|
| 194 |
+
"private_evaluation_arrays_included": boolean
|
| 195 |
+
```
|
| 196 |
+
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/README.md
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
Predict downstream measured RNA responses to held-out single-gene CRISPRi interventions in K562. All NPZ files are numeric/Unicode and must be loaded without pickle. See metadata.json for normalization, ordering, score and resource contracts. train_condition_stats.npz contains training-only means, per-cell sample variances and cell counts; validation_targets.npz contains measured validation GT-half means and scoring weights for tuning; test_queries.json contains identities only. All target Ensembl IDs occur as readout rows in genes.json. Their observed expression across training conditions is public information; their held-out intervention responses are not supplied. The directly targeted gene has zero scoring weight for its own condition. Training and test conditions are disjoint by canonical gene, and technical cell halves are not biological replicates.
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/genes.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/manifest.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema_version": 1,
|
| 3 |
+
"files": {
|
| 4 |
+
"README.md": {
|
| 5 |
+
"sha256": "517dc558584be77eb2549d791a9784028c061a6e2998abea72897eee1c7e7b77",
|
| 6 |
+
"bytes": 868
|
| 7 |
+
},
|
| 8 |
+
"genes.json": {
|
| 9 |
+
"sha256": "b2b436ab503e8eb80d291bfc527dbad7777ecb27a7cf6806cad96fed1708559a",
|
| 10 |
+
"bytes": 153793
|
| 11 |
+
},
|
| 12 |
+
"metadata.json": {
|
| 13 |
+
"sha256": "6541388b6ad8270b791c909c3c09d1a548108838b9f0f969fef7ded7861cc04b",
|
| 14 |
+
"bytes": 3778
|
| 15 |
+
},
|
| 16 |
+
"provenance.json": {
|
| 17 |
+
"sha256": "8ede8971696cd9b9b58ae5767de4b64cc3c830e3012dd814f20b0647230f576e",
|
| 18 |
+
"bytes": 1905
|
| 19 |
+
},
|
| 20 |
+
"test_queries.json": {
|
| 21 |
+
"sha256": "c6142c4a549256e30994a28e1747b3ad1a76ac4c6a72ab7474138de2cb3d067a",
|
| 22 |
+
"bytes": 5423
|
| 23 |
+
},
|
| 24 |
+
"train_condition_stats.npz": {
|
| 25 |
+
"sha256": "213a9cfadf6122065a5f5f5beec43efd18c4ec70015e42e61d142a78fe5c2ae4",
|
| 26 |
+
"bytes": 54440658
|
| 27 |
+
},
|
| 28 |
+
"validation_targets.npz": {
|
| 29 |
+
"sha256": "d40483f6f1f22c35faa091d3150e163655827939fc53366565e948b72a872a72",
|
| 30 |
+
"bytes": 11697141
|
| 31 |
+
}
|
| 32 |
+
},
|
| 33 |
+
"public_test_content": "233canonicaltargetEnsemblIDs and requiredshape only; allreadoutgeneidentities",
|
| 34 |
+
"private_evaluation_arrays_included": false
|
| 35 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/metadata.json
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"dataset": "ReplogleWeissman2022_K562_essential",
|
| 3 |
+
"expression": "Mean of per-cell log1p(10000*RNAcounts/sum(all8563verifiedreleasedRNA))",
|
| 4 |
+
"target_semantics": "Unseen single canonical target ENSG; all guides/cells of each target assigned together",
|
| 5 |
+
"train_stats": "control row then1082trainingconditions; variance is per-cell sample variance (ddof1), n_cells counts cells; means have not used heldout perturbation responses",
|
| 6 |
+
"genes": "All output and target descriptors share exact genes.json gene_ids order; all1546eligibletargets are readout rows",
|
| 7 |
+
"validation": "Only validation GT-half means and GT-half-derived downstream metricweights are public for tuning; noTD",
|
| 8 |
+
"test": "Only233targetIDs and requested shape; no responses, technicalduplicate, weights, originalvarresponse summaries or uns",
|
| 9 |
+
"primary_metric": "Mean across conditions of sum_g(weight_g*(prediction_g-GT_g)^2); own targetedgene weight=0 and other weights sum1; weights derived from squared minmax abs Welch t versus split-controlGT cells",
|
| 10 |
+
"no_test_response_in_gene_descriptors": true,
|
| 11 |
+
"files": [
|
| 12 |
+
"README.md",
|
| 13 |
+
"genes.json",
|
| 14 |
+
"manifest.json",
|
| 15 |
+
"metadata.json",
|
| 16 |
+
"provenance.json",
|
| 17 |
+
"test_queries.json",
|
| 18 |
+
"train_condition_stats.npz",
|
| 19 |
+
"validation_targets.npz"
|
| 20 |
+
],
|
| 21 |
+
"output_shape": [
|
| 22 |
+
233,
|
| 23 |
+
3323
|
| 24 |
+
],
|
| 25 |
+
"task_id": "biology.unseen_perturbation_rna_response",
|
| 26 |
+
"schema_version": 1,
|
| 27 |
+
"cell_line": "K562",
|
| 28 |
+
"perturbation_type": "single-target CRISPR interference",
|
| 29 |
+
"normalization_denominator_features": 8563,
|
| 30 |
+
"feature_selection": "2000 highly variable RNA genes chosen using training target/control cells only, union all1546eligible targetRNA genes;3323total in original RNA-column order",
|
| 31 |
+
"scientific_scoring": {
|
| 32 |
+
"primary_loss": "mean over233conditions of sum over3323genes w*(prediction-measuredGTmean)^2",
|
| 33 |
+
"own_target_weight": "Exactly0 for the canonical target ENSG; remaining weights renormalizedsum1",
|
| 34 |
+
"weight_construction": "GT-half versus same-role controlGT-half Welch t with denominator sqrt(sample_variance/n+control_sample_variance/n); zero denominator t=0; minmax abs(t) over all3323genes, square, then self=0 and renormalize; zero remaining weights are an author data error",
|
| 35 |
+
"score_formula": "clip(100*(B-L)/(B-T),0,100)",
|
| 36 |
+
"B": 0.03441638519456471,
|
| 37 |
+
"T": 0.00714613515111071,
|
| 38 |
+
"B_definition": "Loss of equal-weight mean response of1082noncontroltrainingconditions",
|
| 39 |
+
"T_definition": "Loss of private technicalsecondhalfmeans against private GThalfmeans under same weights",
|
| 40 |
+
"GT_origin": "Private measured cell statistics, not an author model prediction",
|
| 41 |
+
"technical_half_limit": "Same experimental cohort, not independent biological replicates or a universal attainable predictor bound"
|
| 42 |
+
},
|
| 43 |
+
"output": {
|
| 44 |
+
"path": "results/predictions.npy",
|
| 45 |
+
"dtype": [
|
| 46 |
+
"float32",
|
| 47 |
+
"float64"
|
| 48 |
+
],
|
| 49 |
+
"shape": [
|
| 50 |
+
233,
|
| 51 |
+
3323
|
| 52 |
+
],
|
| 53 |
+
"finite": true,
|
| 54 |
+
"maximum_bytes": 8388608,
|
| 55 |
+
"source_path": "analysis.py",
|
| 56 |
+
"maximum_source_bytes": 65536,
|
| 57 |
+
"maximum_compressed_source_bytes": 12288,
|
| 58 |
+
"maximum_source_literal_characters": 8192,
|
| 59 |
+
"maximum_single_literal_characters": 1024,
|
| 60 |
+
"literal_definition": "Characters of all string, bytes, f-string, numeric and comment tokens in analysis.py; compressed size is zlib level 9 of the source bytes"
|
| 61 |
+
},
|
| 62 |
+
"resource_contract": {
|
| 63 |
+
"fresh_replay_seconds": 1200,
|
| 64 |
+
"CPUs": 4,
|
| 65 |
+
"threads": 4,
|
| 66 |
+
"memory_GiB": 8,
|
| 67 |
+
"development_budget": "Read timeout_seconds in task_info.json; configured independently of replay",
|
| 68 |
+
"scientific_data_access": "Only supplied public data; no external datasets, models or knowledge resources; provider connectivity is separate",
|
| 69 |
+
"replay_atol": 1e-05,
|
| 70 |
+
"replay_rtol": 1e-06
|
| 71 |
+
}
|
| 72 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/provenance.json
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema_version": 1,
|
| 3 |
+
"dataset": "Replogle et al. (2022) genome-scale Perturb-seq, K562 essential-gene CRISPRi, day 6",
|
| 4 |
+
"license": "CC-BY-4.0",
|
| 5 |
+
"original_paper": "https://doi.org/10.1016/j.cell.2022.05.013",
|
| 6 |
+
"feature_identity_audit": "All8563releasedfeatureEnsemblIDs uniquely match original Gene Expression annotation;0guide/nonRNA/engineeredcolumns retained. Original CellRanger table includes36601RNA+4561guidefeatures; releasedX was checked explicitly.",
|
| 7 |
+
"source_prefilter": "Released raw counts already restrict genes to mean>0.01UMI/cell. Our normalization sums all8563verifiedreleasedRNAgenes before panel selection.",
|
| 8 |
+
"target_identity": "obs.gene_id exact var.ensembl_id mapping;TBCE and HSPA14 resolved by ID, not symbol guessing.",
|
| 9 |
+
"eligible_cohort": "1546 target IDs with >=64 cells and unique measured RNA identity",
|
| 10 |
+
"split": "Fixed gene-disjoint assignment of canonical target IDs to training (1082), validation (231) and test (233); all guides/cells per target stay together. The assignment recipe is private.",
|
| 11 |
+
"technical_halves": "Each condition was divided once into two fixed technical cell halves; the first supplies GT statistics. The division recipe is private.",
|
| 12 |
+
"training_cells": 188047,
|
| 13 |
+
"training_controls": 7483,
|
| 14 |
+
"validation_controls": 1604,
|
| 15 |
+
"test_controls": 1604,
|
| 16 |
+
"public_numeric_source_hashes": {
|
| 17 |
+
"train_condition_stats.npz": "213a9cfadf6122065a5f5f5beec43efd18c4ec70015e42e61d142a78fe5c2ae4",
|
| 18 |
+
"validation_targets.npz": "d40483f6f1f22c35faa091d3150e163655827939fc53366565e948b72a872a72"
|
| 19 |
+
},
|
| 20 |
+
"offline_boundary": "Use only the supplied public files. Retrieving the original measurements or reconstructing the held-out split from external data breaches the declared holdout.",
|
| 21 |
+
"no_original_full_data_summary_priors": "Original var mean/std/cv/fano/ncounts/ncells and obs normalization summaries are not exported as predictive inputs."
|
| 22 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/test_queries.json
ADDED
|
@@ -0,0 +1,241 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"condition_ids": [
|
| 3 |
+
"ENSG00000188846",
|
| 4 |
+
"ENSG00000168028",
|
| 5 |
+
"ENSG00000075856",
|
| 6 |
+
"ENSG00000108654",
|
| 7 |
+
"ENSG00000149636",
|
| 8 |
+
"ENSG00000198700",
|
| 9 |
+
"ENSG00000129055",
|
| 10 |
+
"ENSG00000112578",
|
| 11 |
+
"ENSG00000088205",
|
| 12 |
+
"ENSG00000166902",
|
| 13 |
+
"ENSG00000171314",
|
| 14 |
+
"ENSG00000124784",
|
| 15 |
+
"ENSG00000075131",
|
| 16 |
+
"ENSG00000134697",
|
| 17 |
+
"ENSG00000153187",
|
| 18 |
+
"ENSG00000089289",
|
| 19 |
+
"ENSG00000143436",
|
| 20 |
+
"ENSG00000101161",
|
| 21 |
+
"ENSG00000101882",
|
| 22 |
+
"ENSG00000196235",
|
| 23 |
+
"ENSG00000108592",
|
| 24 |
+
"ENSG00000070785",
|
| 25 |
+
"ENSG00000169756",
|
| 26 |
+
"ENSG00000090889",
|
| 27 |
+
"ENSG00000139618",
|
| 28 |
+
"ENSG00000100316",
|
| 29 |
+
"ENSG00000204394",
|
| 30 |
+
"ENSG00000198755",
|
| 31 |
+
"ENSG00000167118",
|
| 32 |
+
"ENSG00000107937",
|
| 33 |
+
"ENSG00000183751",
|
| 34 |
+
"ENSG00000141030",
|
| 35 |
+
"ENSG00000094914",
|
| 36 |
+
"ENSG00000126814",
|
| 37 |
+
"ENSG00000004779",
|
| 38 |
+
"ENSG00000067334",
|
| 39 |
+
"ENSG00000131351",
|
| 40 |
+
"ENSG00000204899",
|
| 41 |
+
"ENSG00000149136",
|
| 42 |
+
"ENSG00000183495",
|
| 43 |
+
"ENSG00000175792",
|
| 44 |
+
"ENSG00000188186",
|
| 45 |
+
"ENSG00000111361",
|
| 46 |
+
"ENSG00000166012",
|
| 47 |
+
"ENSG00000173120",
|
| 48 |
+
"ENSG00000129197",
|
| 49 |
+
"ENSG00000196531",
|
| 50 |
+
"ENSG00000173145",
|
| 51 |
+
"ENSG00000184983",
|
| 52 |
+
"ENSG00000103502",
|
| 53 |
+
"ENSG00000127564",
|
| 54 |
+
"ENSG00000132341",
|
| 55 |
+
"ENSG00000175110",
|
| 56 |
+
"ENSG00000151849",
|
| 57 |
+
"ENSG00000125952",
|
| 58 |
+
"ENSG00000126226",
|
| 59 |
+
"ENSG00000177954",
|
| 60 |
+
"ENSG00000172530",
|
| 61 |
+
"ENSG00000239900",
|
| 62 |
+
"ENSG00000173085",
|
| 63 |
+
"ENSG00000185024",
|
| 64 |
+
"ENSG00000183258",
|
| 65 |
+
"ENSG00000213551",
|
| 66 |
+
"ENSG00000197265",
|
| 67 |
+
"ENSG00000136930",
|
| 68 |
+
"ENSG00000127337",
|
| 69 |
+
"ENSG00000156876",
|
| 70 |
+
"ENSG00000119414",
|
| 71 |
+
"ENSG00000130713",
|
| 72 |
+
"ENSG00000121022",
|
| 73 |
+
"ENSG00000164815",
|
| 74 |
+
"ENSG00000173933",
|
| 75 |
+
"ENSG00000168818",
|
| 76 |
+
"ENSG00000116670",
|
| 77 |
+
"ENSG00000162521",
|
| 78 |
+
"ENSG00000117682",
|
| 79 |
+
"ENSG00000169217",
|
| 80 |
+
"ENSG00000134597",
|
| 81 |
+
"ENSG00000197713",
|
| 82 |
+
"ENSG00000113643",
|
| 83 |
+
"ENSG00000062822",
|
| 84 |
+
"ENSG00000158615",
|
| 85 |
+
"ENSG00000119720",
|
| 86 |
+
"ENSG00000173726",
|
| 87 |
+
"ENSG00000134153",
|
| 88 |
+
"ENSG00000142230",
|
| 89 |
+
"ENSG00000143774",
|
| 90 |
+
"ENSG00000160294",
|
| 91 |
+
"ENSG00000092010",
|
| 92 |
+
"ENSG00000108384",
|
| 93 |
+
"ENSG00000136908",
|
| 94 |
+
"ENSG00000142784",
|
| 95 |
+
"ENSG00000108506",
|
| 96 |
+
"ENSG00000135775",
|
| 97 |
+
"ENSG00000196943",
|
| 98 |
+
"ENSG00000172171",
|
| 99 |
+
"ENSG00000159259",
|
| 100 |
+
"ENSG00000126883",
|
| 101 |
+
"ENSG00000125107",
|
| 102 |
+
"ENSG00000150753",
|
| 103 |
+
"ENSG00000132912",
|
| 104 |
+
"ENSG00000130725",
|
| 105 |
+
"ENSG00000063322",
|
| 106 |
+
"ENSG00000137411",
|
| 107 |
+
"ENSG00000108883",
|
| 108 |
+
"ENSG00000115207",
|
| 109 |
+
"ENSG00000163468",
|
| 110 |
+
"ENSG00000162368",
|
| 111 |
+
"ENSG00000146909",
|
| 112 |
+
"ENSG00000136492",
|
| 113 |
+
"ENSG00000090273",
|
| 114 |
+
"ENSG00000073111",
|
| 115 |
+
"ENSG00000161970",
|
| 116 |
+
"ENSG00000021776",
|
| 117 |
+
"ENSG00000157778",
|
| 118 |
+
"ENSG00000099795",
|
| 119 |
+
"ENSG00000106628",
|
| 120 |
+
"ENSG00000101444",
|
| 121 |
+
"ENSG00000110025",
|
| 122 |
+
"ENSG00000108344",
|
| 123 |
+
"ENSG00000123297",
|
| 124 |
+
"ENSG00000167842",
|
| 125 |
+
"ENSG00000163930",
|
| 126 |
+
"ENSG00000127554",
|
| 127 |
+
"ENSG00000167113",
|
| 128 |
+
"ENSG00000184260",
|
| 129 |
+
"ENSG00000116221",
|
| 130 |
+
"ENSG00000044574",
|
| 131 |
+
"ENSG00000008128",
|
| 132 |
+
"ENSG00000168538",
|
| 133 |
+
"ENSG00000163029",
|
| 134 |
+
"ENSG00000114573",
|
| 135 |
+
"ENSG00000106803",
|
| 136 |
+
"ENSG00000088325",
|
| 137 |
+
"ENSG00000135521",
|
| 138 |
+
"ENSG00000093009",
|
| 139 |
+
"ENSG00000136888",
|
| 140 |
+
"ENSG00000163655",
|
| 141 |
+
"ENSG00000143319",
|
| 142 |
+
"ENSG00000104976",
|
| 143 |
+
"ENSG00000132664",
|
| 144 |
+
"ENSG00000102030",
|
| 145 |
+
"ENSG00000157349",
|
| 146 |
+
"ENSG00000174748",
|
| 147 |
+
"ENSG00000112096",
|
| 148 |
+
"ENSG00000278053",
|
| 149 |
+
"ENSG00000138071",
|
| 150 |
+
"ENSG00000100836",
|
| 151 |
+
"ENSG00000155115",
|
| 152 |
+
"ENSG00000090470",
|
| 153 |
+
"ENSG00000080986",
|
| 154 |
+
"ENSG00000133226",
|
| 155 |
+
"ENSG00000108651",
|
| 156 |
+
"ENSG00000131558",
|
| 157 |
+
"ENSG00000165630",
|
| 158 |
+
"ENSG00000119787",
|
| 159 |
+
"ENSG00000196504",
|
| 160 |
+
"ENSG00000159055",
|
| 161 |
+
"ENSG00000159147",
|
| 162 |
+
"ENSG00000196449",
|
| 163 |
+
"ENSG00000164609",
|
| 164 |
+
"ENSG00000091651",
|
| 165 |
+
"ENSG00000014123",
|
| 166 |
+
"ENSG00000136238",
|
| 167 |
+
"ENSG00000115484",
|
| 168 |
+
"ENSG00000198793",
|
| 169 |
+
"ENSG00000211456",
|
| 170 |
+
"ENSG00000249115",
|
| 171 |
+
"ENSG00000186591",
|
| 172 |
+
"ENSG00000061936",
|
| 173 |
+
"ENSG00000183770",
|
| 174 |
+
"ENSG00000144029",
|
| 175 |
+
"ENSG00000119729",
|
| 176 |
+
"ENSG00000184634",
|
| 177 |
+
"ENSG00000136271",
|
| 178 |
+
"ENSG00000100142",
|
| 179 |
+
"ENSG00000008018",
|
| 180 |
+
"ENSG00000103356",
|
| 181 |
+
"ENSG00000124207",
|
| 182 |
+
"ENSG00000108064",
|
| 183 |
+
"ENSG00000176444",
|
| 184 |
+
"ENSG00000083093",
|
| 185 |
+
"ENSG00000140105",
|
| 186 |
+
"ENSG00000010256",
|
| 187 |
+
"ENSG00000127463",
|
| 188 |
+
"ENSG00000167182",
|
| 189 |
+
"ENSG00000139168",
|
| 190 |
+
"ENSG00000174173",
|
| 191 |
+
"ENSG00000119718",
|
| 192 |
+
"ENSG00000239779",
|
| 193 |
+
"ENSG00000130811",
|
| 194 |
+
"ENSG00000149923",
|
| 195 |
+
"ENSG00000088766",
|
| 196 |
+
"ENSG00000152404",
|
| 197 |
+
"ENSG00000131368",
|
| 198 |
+
"ENSG00000231500",
|
| 199 |
+
"ENSG00000115540",
|
| 200 |
+
"ENSG00000163832",
|
| 201 |
+
"ENSG00000113013",
|
| 202 |
+
"ENSG00000141367",
|
| 203 |
+
"ENSG00000134480",
|
| 204 |
+
"ENSG00000100413",
|
| 205 |
+
"ENSG00000010322",
|
| 206 |
+
"ENSG00000181610",
|
| 207 |
+
"ENSG00000112159",
|
| 208 |
+
"ENSG00000168827",
|
| 209 |
+
"ENSG00000101057",
|
| 210 |
+
"ENSG00000149809",
|
| 211 |
+
"ENSG00000170445",
|
| 212 |
+
"ENSG00000107951",
|
| 213 |
+
"ENSG00000137413",
|
| 214 |
+
"ENSG00000198056",
|
| 215 |
+
"ENSG00000243927",
|
| 216 |
+
"ENSG00000105176",
|
| 217 |
+
"ENSG00000135913",
|
| 218 |
+
"ENSG00000180992",
|
| 219 |
+
"ENSG00000072501",
|
| 220 |
+
"ENSG00000142864",
|
| 221 |
+
"ENSG00000165417",
|
| 222 |
+
"ENSG00000188486",
|
| 223 |
+
"ENSG00000060688",
|
| 224 |
+
"ENSG00000213465",
|
| 225 |
+
"ENSG00000104824",
|
| 226 |
+
"ENSG00000172053",
|
| 227 |
+
"ENSG00000112658",
|
| 228 |
+
"ENSG00000070495",
|
| 229 |
+
"ENSG00000161057",
|
| 230 |
+
"ENSG00000164109",
|
| 231 |
+
"ENSG00000165684",
|
| 232 |
+
"ENSG00000176108",
|
| 233 |
+
"ENSG00000141759",
|
| 234 |
+
"ENSG00000204152",
|
| 235 |
+
"ENSG00000165501"
|
| 236 |
+
],
|
| 237 |
+
"shape": [
|
| 238 |
+
233,
|
| 239 |
+
3323
|
| 240 |
+
]
|
| 241 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/train_condition_stats.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:213a9cfadf6122065a5f5f5beec43efd18c4ec70015e42e61d142a78fe5c2ae4
|
| 3 |
+
size 54440658
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/data/validation_targets.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d40483f6f1f22c35faa091d3150e163655827939fc53366565e948b72a872a72
|
| 3 |
+
size 11697141
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/reference.json
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task_id": "biology.unseen_perturbation_rna_response",
|
| 3 |
+
"schema_version": 1,
|
| 4 |
+
"shape": [
|
| 5 |
+
233,
|
| 6 |
+
3323
|
| 7 |
+
],
|
| 8 |
+
"B": 0.03441638519456471,
|
| 9 |
+
"T": 0.00714613515111071,
|
| 10 |
+
"public_input_sha256": {
|
| 11 |
+
"README.md": "517dc558584be77eb2549d791a9784028c061a6e2998abea72897eee1c7e7b77",
|
| 12 |
+
"genes.json": "b2b436ab503e8eb80d291bfc527dbad7777ecb27a7cf6806cad96fed1708559a",
|
| 13 |
+
"manifest.json": "f5cfa2ac4af1c21dccf1d031f775dadf8c4e68459c805bddf2b7f62eef183c88",
|
| 14 |
+
"metadata.json": "6541388b6ad8270b791c909c3c09d1a548108838b9f0f969fef7ded7861cc04b",
|
| 15 |
+
"provenance.json": "8ede8971696cd9b9b58ae5767de4b64cc3c830e3012dd814f20b0647230f576e",
|
| 16 |
+
"test_queries.json": "c6142c4a549256e30994a28e1747b3ad1a76ac4c6a72ab7474138de2cb3d067a",
|
| 17 |
+
"train_condition_stats.npz": "213a9cfadf6122065a5f5f5beec43efd18c4ec70015e42e61d142a78fe5c2ae4",
|
| 18 |
+
"validation_targets.npz": "d40483f6f1f22c35faa091d3150e163655827939fc53366565e948b72a872a72"
|
| 19 |
+
},
|
| 20 |
+
"reference_npz_sha256": "24da3819f2943905ce53b1995e8946e978673626ccc21a026e06a34ded36d123",
|
| 21 |
+
"oracle_output_sha256": "bc1365944c0dccc7310ad5ddbfee64b8f3acb0e0e7a232e618e879a79d148732",
|
| 22 |
+
"reference_kind": "private measured GT-half statistics; no predictive author model",
|
| 23 |
+
"scientific_split": "fixed canonical gene split seed42; all integer harness seeds yield same scientific instance",
|
| 24 |
+
"primary_scope": "Downstream RNA; exact own target Ensembl column has zero weight",
|
| 25 |
+
"replay_gate_GT_selfcheck": "Only prediction_replay_consistency skipped; scientific score still checked"
|
| 26 |
+
}
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/reference.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:24da3819f2943905ce53b1995e8946e978673626ccc21a026e06a34ded36d123
|
| 3 |
+
size 17734080
|
tasks/biology.unseen_perturbation_rna_response__seed31415/reference/results/predictions.npy
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bc1365944c0dccc7310ad5ddbfee64b8f3acb0e0e7a232e618e879a79d148732
|
| 3 |
+
size 6194200
|
tasks/biostat.adaptive_cohort_transport__seed31415/data/experiments.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tasks/biostat.adaptive_cohort_transport__seed31415/data/policies.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|